CUTS: A Deep Learning and Topological

Framework for Multigranular

Unsupervised Medical Image Segmentation

Chen Liu1∗ Matthew Amodio1∗ Liangbo L. Shen2 Feng Gao3 Arman Avesta4
Sanjay Aneja4§ Jay C. Wang5,6§ Lucian V. Del Priore5§ Smita Krishnaswamy1,3§
arXiv:2209.11359v7 [[Link]] 25 Jun 2024

1
Yale University Department of Computer Science
2
University of California, San Francisco, Department of Ophthalmology
3
Yale University Department of Genetics 4 Yale University Department of Therapeutic Radiology
5
Yale University Department of Ophthalmology 6 Northern California Retina Vitreous Associates
* These authors are joint first authors. § Senior authors.
Please direct correspondence to: [Link]@[Link] or [Link]@[Link].

Abstract
Segmenting medical images is critical to facilitating both patient diagnoses and
quantitative research. A major limiting factor is the lack of labeled data, as ob-
taining expert annotations for each new set of imaging data and task can be labor
intensive and inconsistent among annotators. We present CUTS, an unsupervised
deep learning framework for medical image segmentation. CUTS operates in two
stages. For each image, it produces an embedding map via intra-image contrastive
learning and local patch reconstruction. Then, these embeddings are partitioned
at dynamic granularity levels that correspond to the data topology. CUTS yields
a series of coarse-to-fine-grained segmentations that highlight features at various
granularities. We applied CUTS to retinal fundus images and two types of brain
MRI images to delineate structures and patterns at different scales. When eval-
uated against predefined anatomical masks, CUTS improved the dice coefficient
and Hausdorff distance by at least 10% compared to existing unsupervised meth-
ods. Finally, CUTS showed performance on par with Segment Anything Models
(SAM, MedSAM, SAM-Med2D) pre-trained on gigantic labeled datasets.
The code is available at [Link]

1 Introduction
Medical image segmentation plays an increasingly crucial role in both research and clinical settings
in a wide array of imaging modalities including microscopy, X-ray, ultrasound, optical coherence to-
mography (OCT), computed tomography (CT), magnetic resonance imaging (MRI), positron emis-
sion tomography (PET), and others [1]. With high-quality image segmentation, clinicians can more
easily diagnose and monitor the progression of diseases to improve patient care. Traditional med-
ical image segmentation methods rely on hand-crafted features [2–7] or predefined atlases [8–10].
These methods are gradually being replaced by deep learning [11–14] as supervised neural networks
demonstrate superior performance than feature-based methods and less overhead than atlas-based
methods. Although supervised neural networks have been widely successful in image segmentation
in recent years, there are several issues in applying them to medical images, particularly in order
to make clinical inferences. First, these networks depend on expert annotations, so they require a
large number of labels to adequately cover the data variance to produce reliable segmentations [12].
Second, supervised networks trained on one set of annotated images can fail to generalize to similar

27th International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI 2024).
images collected in very slightly different contexts, such as in different patient populations or on
different devices [15]. Third, the desired segmentation granularity may vary across use cases even
if the exact same image is concerned — for example, localizing a brain tumor would require a finer
segmentation compared to measuring the brain volume — yet this need is not easily accommodated
by supervised approaches without updating the labels.
To address these issues, we propose to automatically segment medical images using an entirely
unsupervised framework that combines recent advances in representation learning with advances
in data geometry and topology. An unsupervised approach circumvents the need for costly expert
annotations and alleviates the cross-domain generalization problem. More importantly, we also
design our approach to produce multigranular segmentations, which can potentially target multiple
regions of interest without supervision.
Our framework, which we denote Contrastive and Unsupervised Training for multigranular medical
image Segmentation (CUTS), was named as an homage to the renowned painter Henri Matisse,
who famously used a “cut-up” method he called “drawing with scissors” to assemble an image
based on patches of material from different sources. Our technique is in essence the reverse of this
process, as we start with the initial image and use unsupervised machine learning to segment the
initial figure into a collection of relatively homogeneous patches using data coarse graining in a
learned latent space. Although it may seem trivial to identify the different pieces of paper cut up by
scissors, segmentation of medical images is more challenging as the boundaries between biological
structures, such as between healthy and pathological tissues, are not always sharp and clean.
CUTS is designed as an unsupervised segmentation pipeline. The images are processed in units
of pixel-centered patches, which consists of a fixed-size crop of image centered on an image pixel.
A convolutional encoder is then trained on these pixel-centered patches with both intra-image con-
trastive learning and local patch reconstruction as optimization objectives. We note that contrastive
patches should come from the domain of the medical image itself to create a meaningful pixel
embedding. Thus, we find suitable contrastive patches within each image itself using an image
similarity metric. Subsequently, the learned embedding space serves as a stronger feature-rich foun-
dation for a multiscale, topology-based data coarse graining method called diffusion condensation
that produces multigranular segmentations.

Our main contributions include:

• CUTS, a novel unsupervised framework with a two-stage approach: it first produces an

image-specific pixel-centered patch embedding via a convolutional encoder, and subse-
quently uses diffusion condensation to coarse-grain these patches into clusters at various
levels of granularity to perform multiscale segmentation.
• (Specific to the first stage) A novel optimization objective that combines intra-image con-
trastive learning with local patch reconstruction to help the convolutional encoder learn
an expressive embedding space.
• (Specific to the second stage) A multiscale cluster assignment approach that utilizes diffu-
sion condensation, which provides clinicians with labels at multiple levels of segmenta-
tion granularities, potentially highlighting clinically relevant regions at various scales.

The remainder of this paper is organized as follows. First, we discuss related work in the field. With
this provided context, we then introduce our framework detailing the neural network architecture,
optimization objective, and multiscale segmentation. Finally, we apply our framework on a series of
medical image datasets consisting of retinal fundus images and brain MRI images We evaluate the
performance of CUTS through qualitative and quantitative metrics and compare to several baselines
including other unsupervised approaches and supervised approaches.

2 Related Works
Traditional methods for medical image segmentation Traditional image segmentation methods
generally fall into two categories. The first category relies on hand-crafted image features, such as
line/edge detection [2], graph cuts [3, 7], active contours [4], watershed [5, 6], level-set [16], and
feature clustering [17]. The methods in this category are simple to execute, but they usually struggle

2
with images with more complicated colors and textures. The second category utilizes a precomputed
and annotated atlas to propagate prior knowledge, by warping a predefined set of labels onto new
images through image registration [8–10]. These methods require building and annotating an atlas
for each image dataset – a time-consuming process that may sometimes be impractical.

Supervised learning for medical image segmentation Supervised deep learning has outper-
formed traditional methods in segmenting medical images in the past decade [13]. In supervised
deep learning, a neural network learns to perform a designated task through a data-driven param-
eter optimization process [11]. In medical image segmentation, the most well-known method is
U-Net [18], followed by a proliferation of variants with skip connections, attention mechanisms,
etc. [19–24]. They are all supervised learning methods and thereby require an abundance of expert
annotations.

Towards unsupervised learning With a growing emphasis on avoiding reliance on human expert
annotations, researchers have been exploring unsupervised learning approaches for medical image
segmentation. Many works focused on training with fewer data [25–29]. SSL-ALPNet [30] pro-
posed to directly learn from pseudo-labels generated from Felzenswalb segmentation [7], thus cate-
gorizing it as an unsupervised learning method despite a supervised learning approach. DCGN [31]
used a constrained Gaussian mixture model to cluster pixel representations in histopathology im-
ages. It assumes that different tissue types correspond to different colors, which is not necessarily
true in many other medical image modalities.
Atlas-based unsupervised learning is another promising direction. Compared to their traditional
counterparts [8–10], the versions empowered by deep learning [32, 33] have improved results. When
the domain gap is small, they can be highly effective; otherwise, these methods could fail similarly.
Given their requirement for spatial registration, they are more suitable for clearly defined structures
that show little variation among individuals and thus are less applicable to image domains with
greater variability.

Contrastive learning Contrastive learning [34] was proposed as a generic self-supervised method
to address the issue of limited annotations. Conceptually, it allows neural networks to learn mean-
ingful representations in the embedding space by encouraging similar image pairs to be embedded
closer to each other and vice versa. After a meaningful embedding space is trained, additional lay-
ers can be attached and fine-tuned for downstream tasks. In particular, commonly used contrastive
learning methods such as SimCLR [34], SwaV [35], MoCo [36], BYOL [37], BarlowTwins [38]
and SimSiam [39] focus on extracting image-level representations with an inter-image contrastive
objective. These image-level contrastive learning methods yield no information about intra-image
features and are therefore unsuitable for tasks that require closer scrutiny within the same image,
such as image segmentation. In an attempt to adapt contrastive learning to tackle the image seg-
mentation task, [40] proposed learning image and patch representations through global and local
contrastive training. In [41], the authors used a similar approach, although they coined different
terminologies. Both methods include a supervised fine-tuning stage after contrastive pre-training,
which still depends on labels.

Unsupervised image segmentation with contrastive learning Two leading unsupervised im-
age segmentation methods, DFC [42] and STEGO [43], both utilize contrastive learning concepts.
STEGO learns feature relationships between an image and itself, its k most similar images, and
dissimilar images. Although STEGO can be trained without labels, it relies on pre-trained vision
backbones for knowledge distillation, which is not a requirement in our method. DFC is by far the
most similar to our approach, yet with two key differences. First, DFC contrasts on pixels, while we
operate on pixel-centered patches. Pixel-centered patches contain significantly richer semantic and
textural information than pixels. Second, we achieve segmentation through a topological multiscale
coarse-graining method that produces many segmentation maps at various granularities rather than
a single segmentation map.

Segment Anything Model (SAM) and medical variants Segment Anything Model (SAM) [44]
recently introduced a general-purpose segmentation tool pre-trained on a gigantic dataset of natural
images. As previous researchers have shown [45], SAM offers an alternative solution to label-free
medical image segmentation through an interface called “zero-shot transfer”, where a single point is

3
provided as a prompt which is deciphered by a prompt encoder and sent to a mask model to produce
a segmentation mask. Alternative input formats, such as text prompt (written text) or box prompt
(bounding box) are also supported by this framework.
To better adapt to medical image applications, researchers have developed counterparts that are
pre-trained on large datasets of medical images instead of natural images. MedSAM [46] and SAM-
Med2D [47] are among the most popular variants.
Strictly speaking, SAM and its variants are not unsupervised learning methods. As a result, they still
face the cross-domain generalization problem as previously mentioned, while their brute-force solu-
tion is to cover the entire data distribution with the huge training set. Despite their non-unsupervised
nature, we decided to include them for comparison, since they are arguably the latest state-of-the-art
segmentation framework.

3 Methods

Figure 1: The CUTS Framework. (A) Overview. (B) Pixel-centered patches are mapped into the
embedding space, jointly optimized by two objectives. (C) Positive and negative patch pairs are
selected based on proximity and structural similarity. (D) Diffusion condensation coarse grains em-
bedding vectors at a series of granularities. (E) Segmentation for any granularity can be performed
by mapping cluster assignments to the image space. Multiscale PHATE (MS-PHATE) [48] is used
for visualization.

The CUTS framework contains two stages (Fig. 1(A)). In the first stage, it encodes each pixel along
with the local neighborhood around it, denoted as a “pixel-centered patch”, into a high-dimensional

4
embedding space by jointly optimizing contrastive learning and autoencoding objectives (Fig. 1(B)).
Unlike most contrastive learning methods that learn from augmented versions of full images, CUTS
learns from regions within the same image (Fig. 1(C)). This emphasizes learning of local, intra-
image features instead of invariance over known image transformations or noise models. This is
especially critical for medical images, since they are globally homogeneous (i.e., images from dif-
ferent participants capture the same body part) yet locally heterogeneous (i.e., nuances in structures
or textures within small areas of the image are essential). In the second stage, these embedding vec-
tors are coarse-grained to many levels of granularity by diffusion condensation [49, 50]. Metastable
granularities can be automatically identified from the condensation homology as granularities with
zero topological activity [50]. Segmentation is performed by assigning labels to pixels that corre-
spond to clusters arising from a particular metastable granularity (Fig. 1(D-E)).

3.1 Learning an embedding space for pixel-centered patches

CUTS uses a convolutional neural network as a patch encoder to map pixel-centered patches from
the image space to a latent embedding space. It has convolution, batch norm, activation but no
pooling – to ensure identical spatial dimension between the image and feature map. Two objectives
are jointly optimized.

Intra-image contrastive loss For any anchor patch Pij ∈ Rp×p×c centered at coordinates (i, j),
+ −
we sample positive patches {Pij } and negative patches {Pij }. Let f denote the convolutional en-
+ +
coder. Anchor embedding zij = f (Pij ), positive embeddings Ω+ := {zij } = {f (Pij )}, and
− − −
negative embeddings Ω := {zij } = {f (Pij )}. After projecting the patches to the latent em-
+
bedding space, we can perform contrastive learning on their respective embedding vectors zij and
−
zij . We mine these positive and negative patches using a combination of a proximity heuristic and
an image similarity metric. Only patches nearby (within ± one patch size) and structurally simi-
lar (SSIM [51] > 0.5) to the anchor patch are considered positive patches. The contrastive loss is
defined by Eqn (1) where sim(·) denotes cosine similarity.

pos X + X −
lcontrast = − log , pos = esim(zij ,zij )/τ , neg = esim(zij ,zij )/τ (1)
neg + −
zij ∈Ω+ zij ∈Ω−

Local patch reconstruction loss In addition to the contrastive loss, we ensure that our embedding
of each pixel-centered patch retains information about the patch around it through a reconstruction
loss. For an embedding zij ∈ Rd , the patch reconstruction loss is lrecon = ||Pij − frecon (zij )||22 ,
where frecon (·) : Rd → Rp×p×c is a patch reconstruction module. In implementation, frecon (·) is
a two-layered fully-connected network with ReLU activation.

Final objective function The final objective function is a weighted sum of the contrastive loss and
reconstruction loss, balanced by a weighting coefficient λ ∈ [0, 1].

loss = λ · lcontrast + (1 − λ) · lrecon (2)

Hyperparameters Three key hyperparameters were tuned empirically (Fig. 2). First, we found
that the optimal patch size for pixel-centered patches is 5 × 5. Then, we determined to sample 8
patches in each image for contrastive learning and reconstruction. Lastly, we set the contrastive loss
coefficient at 0.0001. Note that lcontrast is still nontrivial after weighing, because the numerical
value of lcontrast is more than 3 orders of magnitude higher than lrecon at convergence.

3.2 Coarse-graining for multiscale segmentation

For each image patch Pij centered at coordinates (i, j), the patch encoder encodes it to zij ∈ Rd . We
can assign them to n different clusters {c1 , c2 , ..., cn } using a clustering algorithm cls(·) : Rd → R.
Then, we can create a label map L ∈ RH×W where Lij = cls(zij ). The label map L will be the end
product of CUTS segmentation. Notably, with diffusion condensation, cls(·) changes throughout
the process, and therefore we can generate a rich set of labels.

5
 Figure 2: Effects of hyperparameters.

Diffusion condensation [49, 50] is a dynamic process that sweeps through various levels of granular-
ities to identify natural groupings of data. It iteratively condenses data points towards their neighbors
through a diffusion process, at a rate defined by the diffusion probability between the points. Unlike
most clustering methods, diffusion condensation constructs a full hierarchy of coarse-to-fine granu-
larities where the number of clusters at each granularity is not arbitrarily set but rather inferred from
the underlying structure of the data.
Formally, from a data matrix X N ×d with N observations (in our case, N = W × H = the number
of pixels in an image) and d features, we can construct the local affinity among each observation
pair (m, n) ∈ {1, ..., N } using a Gaussian kernel

||xm −xn ||2

K(xm , xn ) = e− ϵ (3)
K is a N × N Gram matrix whose (m, n) entry is denoted K(xm , xn ) to emphasize the dependency
on the data matrix X. xm and xn are both of dimension Rd . The bandwidth parameter ϵ controls
neighborhood sizes.
Given this affinity matrix K, the diffusion operator is defined by Eqn (4a) where D is the diagonal
degree matrix as shown in Eqn (4b).
X
P = D−1 K (4a) D(xm , xm ) = K(xm , xn ) (4b)
n

The diffusion operator P defines the single-step transition probabilities for a diffusion process over
the data, which can be viewed as a Markovian random walk. To perform multi-step diffusion, one
way is to simulate a time-homogeneous diffusion process by raising the diffusion operator to a
power of t which leads to Xt = Pt X [52]. On the other hand, as shown in [50], we could simulate a
time-inhomogeneous diffusion process by iteratively computing the diffusion operator and the data
matrix in the following manner.

X0 ← X
for t ∈ [1, ..., T ] :
Kt−1 ← K(Xt−1 ) /* using Eq. (3) */
Dt−1 ← D(Kt−1 ) /* using Eq. (4b) */ (5)
Pt−1 ← D−1
t−1 Kt−1 /* using Eq. (4a) */
Xt ← Pt−1 Xt−1

The process of diffusion condensation can be summarized as the alternation between the following
two steps:
1. Computing a time-inhomogeneous diffusion operator from the data at iteration t.
2. Applying this operator to the data, moving points towards the local center of gravity, which
forms the data in iteration t + 1.

6
More details on diffusion condensation can be found in [50]. In this paper, we used the official
implementation ([Link]
We can identify the segments that occur consistently over the series of segmentations, called per-
sistent structures. The terminology “persistence” is a measure defined in diffusion condensation
as clusters that stay separated over many iterations. The discovery of persistent structures can be
achieved by rank-ordering different segments based on their persistence levels, which is quantified
by the number of consecutive diffusion iterations in which the segment stays intact and refrains from
being merged into another segment.
For binary segmentation, we need to convert the multi-class label maps to binary segmentation
masks. Following standard practices [43, 45], we use the ground truth segmentation mask to provide
a hint on how to select the foreground for each image. Specifically, we iterate over each foreground
pixel in the ground truth mask and find the most frequently associated cluster of the corresponding
embedding vector. Then we set all pixels whose embeddings match that cluster label as the fore-
ground. This process effectively finds the most probable cluster label if a pixel is randomly selected
from the foreground region of the ground truth and thus is objective and unbiased.

4 Empirical Results

We prepared three medical image datasets to evaluate our proposed framework. The datasets are
chosen to demonstrate the breadth of applications, as they cover variation in color channels (e.g.,
RGB versus intensity-only), imaging sequences (e.g., T1 versus T2 FLAIR), and organs of interest
(e.g., eye versus brain).

Retinal fundus images We used retinal color fundus images of eyes with Geographic Atrophy
(GA) in the age-related eye disease study group [53, 54]. GA regions were segmented by two graders
and reviewed by a retinal specialist, resulting in 56 retinal images with accurate segmentations.

Brain MRI images (ventricles) We used MRIs of patients from the Alzheimer’s Disease Neu-
roimaging Initiative study [55]. A radiologist manually segmented the brain ventricles on 100 T1-
weighted brain MRIs for our study.

Brain MRI images (tumor) We used MRIs of patients with glioma that were scanned by several
healthcare facilities. Tumor regions of 200 fluid-attenuated inversion recovery (FLAIR) brain MRIs
are segmented by trained medical students and finalized by a board-certified attending neuroradiol-
ogist.

4.1 Qualitative results on multigranular segmentation

As shown in Fig. 3, our multiscale segmentation method provides delineation of image structures
at various granularities. The diffusion condensation process starts when all pixels are isolated from
each other (pure noise, not shown in the figure). After a few iterations, fine-grained structures begin
to emerge, as the most similar pixels are clustered together (leftmost columns starting from the third
column). On these finest scales, even the smallest structures are delineated, such as the retinal ves-
sels in the retinal images (first row). Moving toward the coarser scales, anatomical structures arise
as tiny patterns collectively form larger groups. Signature structures include the optic disc and geo-
graphic atrophy in the retinal images (first row), white and gray matter in the brain ventricles images
(third row), and tumor region in the brain tumor images (fifth row). Detection of these anatomical
structures can facilitate automatic measurements of their sizes, shapes, and locations for clinical
interventions. On the coarser side of the spectrum, most structures are iteratively merged through
diffusion condensation, leaving only the most distinctive objects in the image. The final resolution
(rightmost column) identifies the two remaining clusters which correspond to the foreground and
background, respectively.
Qualitatively, we show that CUTS is able to automatically detect meaningful structures and patterns
at multiple granularities within medical images of various modalities. It enables users to deter-
mine their desired level of detail without the necessity of manually annotating data for the model’s
training.

7
Figure 3: Multigranular segmentation (odd rows) captures distinctive patterns at various scales.
Multiscale PHATE (even rows) is used to visualize the diffusion condensation process. The results
of CUTS + spectral k-means clustering (“k-means”) and CUTS + diffusion condensation persistent
structures (“diffusion-P”) are also shown for reference.

4.2 Qualitative and quantitative results on binary segmentation

We compared the performance of CUTS on the three datasets with several alternative methods.
We first compared it with three traditional unsupervised methods: Otsu’s watershed [5], Felzen-
szwalb [7], and SLIC [17]. We then compared with DFC [42] and STEGO [43], two recent unsuper-
vised models based on deep learning. For each experiment, we re-trained DFC, STEGO, and CUTS
on the images only.
Next, we compared against the Segment Anything Model (SAM) [44, 45] which was pre-trained on
11 million images and 1.1 billion masks, as well as its medical-image variants (MedSAM [46]/SAM-
Med2D [47]) pre-trained on 1.6/4.6 million images and 1.6/19.7 million masks, respectively. For
SAM variants, we provided a center point of the ground truth label as a prompt for segmentation of
each image [45].
Lastly, for reference, we benchmarked a random labeler as the performance lower bound and two
fully supervised methods, UNet [18] and nn-UNet [24], as the upper bound. For coarse-graining of
the pixel embeddings, we also implemented a spectral k-means clustering [56] alternative, which
segments at only one granularity level. For a fair comparison, we applied the same binarization
approach described in the Methods section to all unsupervised methods.

Geographic atrophy segmentation in retinal fundus images Our first experiment aims to seg-
ment regions of geographic atrophy (GA) in retinal fundus images. GA is an advanced stage of
age-related macular degeneration (AMD) characterized by progressive macula degeneration. CUTS
accurately selects the region of atrophy.
Qualitatively, CUTS is better at delineating the boundaries of atrophy compared to all other unsu-
pervised methods (Fig. 4). The quantitative results (Table 1) also confirmed this observation. CUTS
created better segmentations than other unsupervised methods, as indicated by a higher dice score
and a lower Hausdorff distance.

Ventricle segmentation in brain MRI images In our next experiment, we tried to segment the
brain ventricles in MRI images of patients at various stages of Alzheimer’s disease. This task is con-

8
Figure 4: Qualitative segmentation comparison. Green curves outline the ground truth labels while
blue or red curves outline the predictions. “diffusion-B”: the best diffusion condensation granularity.
“Sup.”: supervised “P.T.”: pre-training. “+bbox”: using bounding boxes instead of points as input;
included for completeness but would be unfair for comparison.

sidered clinically important because the volume of the brain ventricles can predict the progression
of dementia [57, 58].
Qualitatively, CUTS delineated the brain ventricles in a wide variety of settings (Fig. 4). Due to the
general trend that ventricles appear consistently darker than the rest of the image, most methods are
able to achieve good overall performance on several cases. However, our method usually delineates
the boundaries better than competing methods, especially for images showing noncontiguous ven-
tricles. The quantitative results (Table 1) also indicate the superior performance of CUTS over other
unsupervised methods.

Tumor segmentation in brain MRI images Our final experiment investigated a different seg-
mentation target in brain MRI images – brain tumors, or more specifically, glioma. Accurate seg-
mentation of tumor areas is crucial for the diagnosis and treatment of brain tumors. This process
can help radiologists provide vital details about the size, position, and form of tumors, which is
important to determine the most appropriate course of clinical care.
Qualitatively, our method demonstrated superior segmentation compared to other unsupervised
methods, as shown in Fig. 4. As a general observation, competing methods struggle to identify
tumors, although they manage to segment the ventricles in a similar imaging modality. This dis-
parity in performance was anticipated, given the pronounced complexity associated with tumor
segmentation compared to ventricles, due to considerably more subtle contrast and morphologi-
cal distinctions. Nevertheless, CUTS overcomes the inherent challenges and successfully segments
tumor regions. CUTS led the other unsupervised methods by a larger margin compared to the less
demanding task of ventricle segmentation.

Comparison with SAM, MedSAM and SAM-Med2D More impressively, as shown in Table 1,
CUTS achieved better results than every SAM variant on at least 2 out of 3 datasets — under fair
comparison of using a single point as input — without relying on billions of annotations.

Ablation study We confirmed that applying diffusion condensation or spectral k-means on the
raw image pixels is suboptimal compared to CUTS (Table 1).

9
Table 1: Quantitative comparisons from 3 random seeds. Among unsupervised methods, the best is
bolded and runner-up is underscored. § Entries with “+bbox” use bounding boxes instead of points
as input. They are included for completeness but would be unfair for comparison. ‡ Diffusion con-
densation will not run since #features = 1 for each pixel in single-channel images. ∗ The suboptimal
performance of MedSAM is expected. According to the authors, “the point prompt is still an exper-
imental function and the model was trained on a small abdomen CT organ segmentation dataset.”
Retinal Atrophy Brain Ventricles Brain Tumor
Deep learning? Topological? DSC ↑ HD ↓ DSC ↑ HD ↓ DSC ↑ HD ↓
Unsupervised, without learning
′
Watershed (IEEE TPAMI 91 [5]) 0.192±0.000 56.32±0.00 0.781±0.000 30.25±0.00 0.073±0.000 95.42±0.00
Felzenszwalb (IJCV′ 04 [7]) 0.592±0.000 27.60±0.00 0.759±0.000 44.80±0.00 0.316±0.000 21.41±0.00
SLIC (IEEE TPAMI′ 12 [17]) 0.567±0.000 28.76±0.00 0.475±0.000 37.96±0.00 0.242±0.000 47.51±0.00
Unsupervised, with learning
DFC (IEEE TIP′ 20 [42]) 0.300±0.020 46.47±1.42 0.631±0.024 34.28±0.57 0.197±0.004 52.51±0.09
STEGO (ICLR′ 22 [43]) 0.649±0.025 34.12±4.06 0.725±0.050 12.59±4.43 0.176±0.104 57.16±14.09
(Ours) CUTS + Spectral k-means 0.675±0.014 26.82±0.88 0.774±0.008 8.31±0.23 0.432±0.010 33.94±0.65
(Ours) CUTS + Diffusion (pers.) 0.604±0.003 21.69±0.44 0.495±0.002 13.36±0.60 0.390±0.004 33.66±0.24
(Ours) CUTS + Diffusion (best) 0.741±0.007 17.76±0.13 0.810±0.006 7.17±0.18 0.486±0.007 25.16±1.12
Ablation: image pixels instead of latent embeddings
Image pixels + Spectral k-means 0.560±0.000 37.97±0.00 0.386±0.000 26.11±0.00 0.240±0.000 51.69±0.00
Image pixels + Diffusion (pers.) 0.405±0.000 61.67±0.00 ‡ ‡ ‡ ‡
Image pixels + Diffusion (best) 0.538±0.000 45.16±0.00 ‡ ‡ ‡ ‡
Lower bound: random label
Random 0.132±0.000 78.45±0.07 0.149±0.000 61.40±0.02 0.057±0.000 95.53±0.02
Upper bound: supervised
SAM (ICCV′ 23 [44], MedIA′ 23 [45]) 0.924±0.000 9.18±0.01 0.644±0.003 30.24±0.19 0.405±0.000 36.14±0.14
SAM-Med2D (ArXiv [47]) 0.548±0.001 14.69±0.00 0.736±0.000 17.38±0.02 0.591±0.001 12.93±0.01
SAM-Med2D+bbox§ 0.882±0.000 5.31±0.00 0.849±0.000 9.78±0.00 0.686±0.000 8.74±0.00
MedSAM∗ (Nat. Commun.′ 24 [46]) 0.079±0.000 32.29±0.02 0.053±0.000 64.00±0.04 0.088±0.001 33.54±0.02
MedSAM+bbox§ 0.889±0.000 5.21±0.00 0.829±0.000 10.60±0.00 0.702±0.000 7.61±0.00
UNet (MICCAI′ 15 [18]) 0.965±0.014 3.78±1.08 0.989±0.001 1.05±0.10 0.867±0.016 8.84±1.10
nnUNet (Nat. Methods′ 21 [24]) 0.937±0.014 6.00±1.35 0.984±0.005 2.10±0.42 0.834±0.024 8.64±1.60

5 Conclusion

CUTS is a deep learning and topological framework that identifies and highlights important medical
image structures using unsupervised learning. Despite the emergence of foundation models, such
as variants of SAM, CUTS remains relevant and insightful. It is lightweight and does not require
extensive annotation and pre-training in large compute warehouses. Additionally, it is clear that
foundation models like SAM necessitate domain-specific fine-tuning for tasks not covered by the
initial supervised pre-training, which highlights the continued relevance of approaches like CUTS
that investigate objectives, modules and techniques to inject the correct inductive biases. Therefore,
CUTS offers a practical and effective alternative in the evolving landscape of medical imaging.

6 Discussion

Current state-of-the-art methods for medical image segmentation are primarily supervised and there-
fore require domain experts to annotate a large number of medical images. Moreover, it is often
infeasible to collect enough images of rare diseases to train supervised learning models. For exam-
ple, in this work we studied a retinal degeneration condition. The number of images available to
any institution of this condition is usually within a hundred, several orders of magnitude fewer than
popular natural image databases for deep learning with millions of images. Furthermore, another
limitation of supervised learning approaches is the domain generalization problem. When a method
is optimized for a specific type of image used for training, its performance may suffer if used on
other types of image, even if they are only slightly dissimilar.
In contrast, unsupervised methods, like CUTS, while more challenging architect, do not require
human expert grading and thereby circumvent this time-consuming, expensive, and labor-intensive
initial step. Unsupervised methods can also be applied to much smaller datasets, ideal for rare
diseases. Unfortunately, prior attempts to use unsupervised learning to segment medical images
have not achieved the desired results. These unsupervised methods often yield subpar performance,
despite having advantages including independence from labels and the ability to generalize to new
datasets while preserving robustness.
CUTS bridges the difficulty of creating unsupervised images by using the key insight that, while
an image as a whole may be hard to segment, pixels forming boundaries of image features may be

10
detectable by their local context. Thus, CUTS features carefully architected losses for local pixel-
centered patch reconstruction and pixel-centered patch-based contrastive losses based on within-
image augmentation of patches. With these unique penalties, CUTS learns an intermediate represen-
tation of a pixel-centered patch embedding for each image. The key advantage of this pixel-centered
patch representation is that it is amenable to not one segmentation, but several multigranular seg-
mentations of the same image via a topological coarse-graining scheme. The final output of CUTS is
thus several segmentations of the image with features of different resolutions of interest for different
clinical queries.
In our brain tumor image dataset, for example, CUTS enables: (1) brain extraction on the coarsest
scale, (2) isolation of white matter, gray matter, and cerebrospinal fluid on the intermediate scale,
and (3) small tumor segmentation on the finest scale. These different features can be important for
different diagnostic purposes such as tumor placement identification for surgical purposes or small
tumor size extraction for the analysis of metastases.
In this work, we demonstrate the application of CUTS to three medical image datasets from differ-
ent medical domains. On the retinal fundus images, the watershed, Felzenszwalb, and DFC focus
primarily on the contours of the retina without distinguishing the geographic atrophy regions, where
the contrast is more subtle. SLIC and STEGO generally perform better, yet they tend to overesti-
mate the region of interest. CUTS avoids all these caveats and consistently segments geographical
atrophy. On brain MRI for ventricle segmentation, the watershed method often ignores the frontal
or lateral half of the ventricles. Felzenszwalb, SLIC, and DFC have difficulty determining the seg-
mentation boundary. STEGO tends to include tissues around the ventricles. CUTS is slightly more
conservative on the boundary regions but nevertheless outperforms other unsupervised methods. On
the brain MRI images for tumor segmentation, the watershed and Felzenszwalb methods merely
isolate the entire brain from the background with no attention to detailed structures. SLIC, DFC,
and STEGO either ignored the tumor region or merged it with the background. CUTS, on the other
hand, is sensitive to tenuous contrast transitions in tumor regions and generates significantly better
segmentations.
In conclusion, CUTS allows us to identify and highlight important medical image structures using an
unsupervised learning approach. This has enormous implications for the expanding field of medical
image evaluation and represents a step forward in identifying important and distinct information
relevant to the clinical interpretation of images. Such interpretation of medical images is crucial
to disease detection in asymptomatic individuals; examples include mammography for screening
for breast cancer, teleophthalmology and automated image analysis for screening diabetes for eye
disease, or vulnerable populations for macular degeneration and glaucoma, and screening of high-
risk populations such as smokers for early lung cancer.

7 Acknowledgements
The authors would like to thank Mengyuan Sun, Aneesha Ahluwalia, Benjamin K. Young, and
Michael M. Park for delineating geographic atrophy borders on fundus photographs.
This work was supported in part by the National Science Foundation (NSF DMS 2327211,
NSF Career Grant 2047856) and the National Institute of Health (NIH 1R01GM130847-01A1,
NIH 1R01GM135929-01).

11
References
[1] Yabo Fu, Yang Lei, Tonghe Wang, Walter J Curran, Tian Liu, and Xiaofeng Yang. A review
of deep learning based methods for medical image multi-organ segmentation. Physica Medica,
85:107–122, 2021.

[2] Pinaki Pratim Acharjya, Ritaban Das, and Dibyendu Ghoshal. Study and comparison of different
edge detectors for image segmentation. Global Journal of Computer Science and Technology,
2012.

[3] Xinjian Chen and Lingjiao Pan. A survey of graph cuts/graph search based medical image
segmentation. IEEE reviews in biomedical engineering, 11:112–124, 2018.

[4] Issam El Naqa, Deshan Yang, Aditya Apte, Divya Khullar, Sasa Mutic, Jie Zheng, Jeffrey D
Bradley, Perry Grigsby, and Joseph O Deasy. Concurrent multimodality image segmentation
by active contours for radiotherapy treatment planning a. Medical physics, 34(12):4738–4749,
2007.

[5] Luc Vincent and Pierre Soille. Watersheds in digital spaces: an efficient algorithm based
on immersion simulations. IEEE Transactions on Pattern Analysis & Machine Intelligence,
13(06):583–598, 1991.

[6] Hyun-Hwa Oh, Kil-Taek Lim, and Sung-Il Chien. An improved binarization algorithm based on
a water flow model for document image with inhomogeneous backgrounds. Pattern Recognition,
38(12):2612–2625, 2005.

[7] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation.
International journal of computer vision, 59(2):167–181, 2004.

[8] Ivana Isgum, Marius Staring, Annemarieke Rutten, Mathias Prokop, Max A Viergever, and
Bram Van Ginneken. Multi-atlas-based segmentation with local decision fusion—application to
cardiac and aortic segmentation in ct scans. IEEE transactions on medical imaging, 28(7):1000–
1010, 2009.

[9] Paul Aljabar, Rolf A Heckemann, Alexander Hammers, Joseph V Hajnal, and Daniel Rueck-
ert. Multi-atlas based segmentation of brain images: atlas selection and its effect on accuracy.
Neuroimage, 46(3):726–738, 2009.

[10] Juan Eugenio Iglesias and Mert R Sabuncu. Multi-atlas segmentation of biomedical images: a
survey. Medical image analysis, 24(1):205–219, 2015.

[11] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. nature, 521(7553):436–
444, 2015.

[12] Andre Esteva, Alexandre Robicquet, Bharath Ramsundar, Volodymyr Kuleshov, Mark De-
Pristo, Katherine Chou, Claire Cui, Greg Corrado, Sebastian Thrun, and Jeff Dean. A guide to
deep learning in healthcare. Nature medicine, 25(1):24–29, 2019.

[13] Intisar Rizwan I Haque and Jeremiah Neubert. Deep learning approaches to biomedical image
segmentation. Informatics in Medicine Unlocked, 18:100297, 2020.

[14] Shervin Minaee, Yuri Y Boykov, Fatih Porikli, Antonio J Plaza, Nasser Kehtarnavaz, and
Demetri Terzopoulos. Image segmentation using deep learning: A survey. IEEE transactions
on pattern analysis and machine intelligence, 2021.

[15] Michael D Abràmoff, Philip T Lavin, Michele Birch, Nilay Shah, and James C Folk. Piv-
otal trial of an autonomous ai-based diagnostic system for detection of diabetic retinopathy in
primary care offices. NPJ digital medicine, 1(1):1–8, 2018.

[16] Andy Tsai, Anthony Yezzi, William Wells, Clare Tempany, Dewey Tucker, Ayres Fan, W Eric
Grimson, and Alan Willsky. A shape-based approach to the segmentation of medical imagery
using level sets. IEEE transactions on medical imaging, 22(2):137–154, 2003.

12
[17] Radhakrishna Achanta, Appu Shaji, Kevin Smith, Aurelien Lucchi, Pascal Fua, and Sabine
Süsstrunk. Slic superpixels compared to state-of-the-art superpixel methods. IEEE transactions
on pattern analysis and machine intelligence, 34(11):2274–2282, 2012.

[18] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for
biomedical image segmentation. In International Conference on Medical image computing and
computer-assisted intervention, pages 234–241. Springer, 2015.

[19] Zongwei Zhou, Md Mahfuzur Rahman Siddiquee, Nima Tajbakhsh, and Jianming Liang.
Unet++: A nested u-net architecture for medical image segmentation. In Deep learning in medi-
cal image analysis and multimodal learning for clinical decision support, pages 3–11. Springer,
2018.

[20] Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa,
Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net:
Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999, 2018.

[21] Özgün Çiçek, Ahmed Abdulkadir, Soeren S Lienkamp, Thomas Brox, and Olaf Ronneberger.
3d u-net: learning dense volumetric segmentation from sparse annotation. In International
conference on medical image computing and computer-assisted intervention, pages 424–432.
Springer, 2016.

[22] Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Led-
sam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A proba-
bilistic u-net for segmentation of ambiguous images. Advances in neural information processing
systems, 31, 2018.

[23] Xin Yu, Qi Yang, Yinchi Zhou, Leon Y Cai, Riqiang Gao, Ho Hin Lee, Thomas Li, Shunxing
Bao, Zhoubing Xu, Thomas A Lasko, et al. Unest: local spatial representation learning with hi-
erarchical transformer for efficient medical segmentation. Medical Image Analysis, 90:102939,
2023.

[24] Fabian Isensee, Paul F Jaeger, Simon AA Kohl, Jens Petersen, and Klaus H Maier-Hein. nnu-
net: a self-configuring method for deep learning-based biomedical image segmentation. Nature
methods, 18(2):203–211, 2021.

[25] Arnab Kumar Mondal, Jose Dolz, and Christian Desrosiers. Few-shot 3d multi-modal medical
image segmentation using generative adversarial learning. arXiv preprint arXiv:1810.12241,
2018.

[26] Cheng Ouyang, Konstantinos Kamnitsas, Carlo Biffi, Jinming Duan, and Daniel Rueckert.
Data efficient unsupervised domain adaptation for cross-modality image segmentation. In Med-
ical Image Computing and Computer Assisted Intervention–MICCAI 2019: 22nd International
Conference, Shenzhen, China, October 13–17, 2019, Proceedings, Part II 22, pages 669–677.
Springer, 2019.

[27] Hanchao Yu, Shanhui Sun, Haichao Yu, Xiao Chen, Honghui Shi, Thomas S Huang, and Ter-
rence Chen. Foal: Fast online adaptive learning for cardiac motion estimation. In Proceedings of
the IEEE/CVF conference on computer vision and pattern recognition, pages 4313–4323, 2020.

[28] Chen Chen, Chen Qin, Huaqi Qiu, Cheng Ouyang, Shuo Wang, Liang Chen, Giacomo Tarroni,
Wenjia Bai, and Daniel Rueckert. Realistic adversarial data augmentation for mr image segmen-
tation. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd
International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part I 23, pages 667–
677. Springer, 2020.

[29] Qi Yang, Xin Yu, Ho Hin Lee, Yucheng Tang, Shunxing Bao, Kristofer S Gravenstein,
Ann Zenobia Moore, Sokratis Makrogiannis, Luigi Ferrucci, and Bennett A Landman. Label
efficient segmentation of single slice thigh ct with two-stage pseudo labels. Journal of Medical
Imaging, 9(5):052405–052405, 2022.

13
[30] Cheng Ouyang, Carlo Biffi, Chen Chen, Turkay Kart, Huaqi Qiu, and Daniel Rueckert. Self-
supervision with superpixels: Training few-shot medical image segmentation without anno-
tation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August
23–28, 2020, Proceedings, Part XXIX 16, pages 762–780. Springer, 2020.
[31] Yang Nan, Peng Tang, Guyue Zhang, Caihong Zeng, Zhihong Liu, Zhifan Gao, Heye Zhang,
and Guang Yang. Unsupervised tissue segmentation via deep constrained gaussian network.
IEEE Transactions on Medical Imaging, 41(12):3799–3811, 2022.
[32] Amy Zhao, Guha Balakrishnan, Fredo Durand, John V Guttag, and Adrian V Dalca. Data
augmentation using learned transformations for one-shot medical image segmentation. In Pro-
ceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8543–
8553, 2019.
[33] Shuxin Wang, Shilei Cao, Dong Wei, Renzhen Wang, Kai Ma, Liansheng Wang, Deyu Meng,
and Yefeng Zheng. Lt-net: Label transfer by learning reversible voxel-wise correspondence
for one-shot medical image segmentation. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition, pages 9162–9171, 2020.
[34] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework
for contrastive learning of visual representations, 2020.
[35] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand
Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances
in Neural Information Processing Systems, 33:9912–9924, 2020.
[36] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for
unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on
computer vision and pattern recognition, pages 9729–9738, 2020.
[37] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre Richemond, Elena
Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar,
et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural
information processing systems, 33:21271–21284, 2020.
[38] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-
supervised learning via redundancy reduction. In International Conference on Machine Learn-
ing, pages 12310–12320. PMLR, 2021.
[39] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Pro-
ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages
15750–15758, 2021.
[40] Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning
of global and local features for medical image segmentation with limited annotations, 2020.
[41] Ke Yan, Jinzheng Cai, Dakai Jin, Shun Miao, Adam P. Harrison, Dazhou Guo, Youbao Tang,
Jing Xiao, Jingjing Lu, and Le Lu. Self-supervised learning of pixel-wise anatomical embed-
dings in radiological images, 2020.
[42] Wonjik Kim, Asako Kanezaki, and Masayuki Tanaka. Unsupervised learning of image seg-
mentation based on differentiable feature clustering. IEEE Transactions on Image Processing,
29:8055–8068, 2020.
[43] Mark Hamilton, Zhoutong Zhang, Bharath Hariharan, Noah Snavely, and William T Freeman.
Unsupervised semantic segmentation by distilling feature correspondences. In International
Conference on Learning Representations, 2022.
[44] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson,
Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In
Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015–4026,
2023.

14
[45] Maciej A Mazurowski, Haoyu Dong, Hanxue Gu, Jichen Yang, Nicholas Konz, and Yixin
Zhang. Segment anything model for medical image analysis: an experimental study. Medical
Image Analysis, page 102918, 2023.
[46] Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in
medical images. Nature Communications, 15(1):654, 2024.
[47] Junlong Cheng, Jin Ye, Zhongying Deng, Jianpin Chen, Tianbin Li, Haoyu Wang, Yanzhou
Su, Ziyan Huang, Jilong Chen, Lei Jiang, et al. Sam-med2d. arXiv preprint arXiv:2308.16184,
2023.
[48] Manik Kuchroo, Jessie Huang, Patrick Wong, Jean-Christophe Grenier, Dennis Shung,
Alexander Tong, Carolina Lucas, Jon Klein, Daniel B Burkhardt, Scott Gigante, et al. Multi-
scale phate identifies multimodal signatures of covid-19. Nature biotechnology, 40(5):681–691,
2022.
[49] Nathan Brugnone, Alex Gonopolskiy, Mark W Moyle, Manik Kuchroo, David van Dijk,
Kevin R Moon, Daniel Colon-Ramos, Guy Wolf, Matthew J Hirn, and Smita Krishnaswamy.
Coarse graining of data via inhomogeneous diffusion condensation. In 2019 IEEE International
Conference on Big Data (Big Data), pages 2624–2633. IEEE, 2019.
[50] Manik Kuchroo, Marcello DiStasio, Eric Song, Eda Calapkulu, Le Zhang, Maryam Ige,
Amar H Sheth, Abdelilah Majdoubi, Madhvi Menon, Alexander Tong, et al. Single-cell anal-
ysis reveals inflammatory interactions driving macular degeneration. Nature Communications,
14(1):2589, 2023.
[51] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssim. In 2010 20th international
conference on pattern recognition, pages 2366–2369. IEEE, 2010.
[52] Ronald R Coifman and Stéphane Lafon. Diffusion maps. Applied and computational harmonic
analysis, 21(1):5–30, 2006.
[53] Matthew D Davis, Ronald E Gangnon, Li Yin Lee, Larry D Hubbard, BE Klein, Ronald Klein,
Frederick L Ferris, Susan B Bressler, Roy C Milton, et al. The age-related eye disease study
severity scale for age-related macular degeneration: Areds report no. 17. Archives of ophthal-
mology (Chicago, Ill.: 1960), 123(11):1484–1498, 2005.
[54] Liangbo L Shen, Mengyuan Sun, Aneesha Ahluwalia, Benjamin K Young, Michael M Park,
Cynthia A Toth, Eleonora M Lad, and Lucian V Del Priore. Relationship of topographic distri-
bution of geographic atrophy to visual acuity in nonexudative age-related macular degeneration.
Ophthalmology Retina, 5(8):761–774, 2021.
[55] Karen L Crawford, Scott C Neu, and Arthur W Toga. The image and data archive at the
laboratory of neuro imaging. Neuroimage, 124:1080–1083, 2016.
[56] Hongyuan Zha, Xiaofeng He, Chris Ding, Ming Gu, and Horst Simon. Spectral relaxation for
k-means clustering. Advances in neural information processing systems, 14, 2001.
[57] Owen T Carmichael, Lewis H Kuller, Oscar L Lopez, Paul M Thompson, Rebecca A Dutton,
Allen Lu, Sharon E Lee, Jessica Y Lee, Howard J Aizenstein, Carolyn Cidis Meltzer, et al.
Ventricular volume and dementia progression in the cardiovascular health study. Neurobiology
of aging, 28(3):389–397, 2007.
[58] Brian R Ott, Ronald A Cohen, Assawin Gongvatana, Ozioma C Okonkwo, Conrad E Johanson,
Edward G Stopa, John E Donahue, Gerald D Silverberg, Alzheimer’s Disease Neuroimaging
Initiative, et al. Brain ventricular volume and cerebrospinal fluid biomarkers of alzheimer’s
disease. Journal of Alzheimer’s disease, 20(2):647–657, 2010.

15