Research Paper (Excerpt – complete short paper)
Title:
A Deep Learning Approach for Early Detection of Lung Cancer from CT Scans
Author:
Zhang Wei
Department of Biomedical Engineering, Shanghai Jiao Tong University
Abstract (approx. 250 words)
Lung cancer remains the leading cause of cancer-related deaths worldwide.
Early diagnosis significantly improves five-year survival rates. However, manual
inspection of computed tomography (CT) scans is time-consuming and prone
to observer variability. This paper proposes a novel deep learning architecture
– LungNet – based on a modified ResNet-50 backbone with channel-wise
attention and multi-scale feature fusion. The model is trained and validated on
the public LIDC-IDRI dataset, which contains 1,018 CT scans annotated by four
radiologists. We achieve a sensitivity of 94.2% and a specificity of 91.7% for
pulmonary nodule detection, outperforming conventional methods such as
SVM (sensitivity 81.5%) and a standard Faster R-CNN (sensitivity 88.3%).
Additionally, we integrate Grad-CAM visualisations to provide explainable
heatmaps, helping clinicians understand which regions influenced the
prediction. The inference time per scan is 0.32 seconds on an NVIDIA RTX 3090
GPU, making the system feasible for real-time clinical deployment. We also
conduct a radiologist study where five experts reviewed 200 cases with and
without LungNet assistance. With the model, the average reading time
decreased from 8.4 to 5.1 minutes per case, while inter-observer agreement
increased (kappa from 0.72 to 0.85). Our results indicate that deep learning can
serve as an effective second reader, reducing false negatives in busy radiology
departments. Future work will focus on multi-modal integration (PET/CT) and
prospective clinical trials.
1. Introduction
Lung cancer is responsible for approximately 1.8 million deaths annually (WHO,
2024). The National Lung Screening Trial demonstrated that low-dose CT
screening reduces mortality by 20% compared to chest X-ray. Despite this
benefit, the widespread adoption of CT generates an enormous volume of
image data. A single CT study contains hundreds of slices, and a radiologist
may need to interpret dozens of such studies daily. Fatigue and distraction
inevitably lead to missed nodules – particularly small ones (≤5 mm) – which are
often early-stage cancers.
Machine learning has been applied to computer-aided detection (CADe) since
the 1990s, first using hand-crafted features (intensity, shape, texture) followed
by a classifier. These traditional CADe systems suffered from high false-positive
rates and poor generalisation across different CT scanners. The advent of deep
convolutional neural networks (CNNs) has revolutionised medical image
analysis. In 2017, the LUNA16 challenge saw CNNs outperform all traditional
methods. However, most existing models focus either on nodule detection or
on benign/malignant classification, rarely both in an end-to-end framework.
Moreover, their black-box nature limits clinical acceptance.
This paper makes the following contributions:
(1) We design LungNet, a single-stage detector that simultaneously localises
nodules and predicts malignancy scores.
(2) We incorporate a channel-wise attention module (squeeze-and-excitation)
to emphasise informative feature maps.
(3) We propose a multi-scale feature fusion strategy that improves small
nodule detection (diameter < 3 mm).
(4) We provide a fully explainable system using Grad-CAM heatmaps.
(5) We evaluate the system not only on standard metrics but also through a
radiologist reading-time study.
2. Materials and Methods
2.1 Dataset
We used the LIDC-IDRI dataset, which includes 1,018 thoracic CT scans from
1,010 patients. Four experienced radiologists independently marked lesions as
“nodule” (≥3 mm), “non-nodule” (<3 mm), or “indeterminate”. Only nodules
with at least three radiologists’ agreement were considered ground truth. After
preprocessing, we obtained 2,631 nodules (diameter range: 3–30 mm). The
dataset was split into training (70%), validation (10%), and test (20%) at the
patient level to avoid data leakage.
2.2 Preprocessing
All CT volumes were resampled to an isotropic voxel spacing of 1 mm × 1 mm
× 1 mm. Hounsfield units were clipped to [-1000, 400] and normalised to zero
mean and unit variance. To handle class imbalance (background vs. nodule), we
used online hard negative mining and a focal loss function.
2.3 Model Architecture – LungNet
LungNet is built upon a pre-trained ResNet-50 as the backbone. We added
three custom components:
Attention block: A squeeze-and-excitation module after each residual block,
with a reduction ratio of 16.
Feature pyramid network (FPN): A top-down pathway with lateral
connections to combine low-level spatial details and high-level semantic
information.
Detection head: Two parallel sub-networks – one for bounding box regression
(smooth L1 loss) and one for classification (focal loss).
The input is a 3D patch of size 96×96×96 mm³. The output includes nodule
coordinates (centre, diameter) and a malignancy probability (0 = benign, 1 =
malignant).
2.4 Training Details
We used the AdamW optimizer with an initial learning rate of 1e-4, weight
decay 1e-5, and a cosine annealing schedule. Batch size was 8 (limited by GPU
memory). Data augmentation included random rotation (±15°), scaling (±10%),
and elastic deformation. The model was trained for 200 epochs, with early
stopping based on validation loss. Implementation used PyTorch 2.0 and
MONAI.
2.5 Radiologist Study
Five board-certified radiologists (3–15 years of experience) participated. They
reviewed 200 test CT cases twice: first unaided, then after a 2-week washout
period with LungNet assistance (heatmaps + nodule suggestions). Reading
time was recorded automatically. Sensitivity, specificity, and Fleiss’ kappa were
computed.
3. Results
3.1 Detection Performance
On the test set (204 scans, 528 nodules), LungNet achieved a sensitivity
of 94.2% (95% CI: 92.1–96.0) at 4 false positives per scan, and a specificity
of 91.7% (95% CI: 89.2–93.8). For small nodules (3–5 mm), sensitivity was
88.5%, a 12-point improvement over the baseline Faster R-CNN. The area
under the ROC curve (AUC) for malignancy classification was 0.93.
3.2 Comparison with Traditional Methods
Method Sensitivity (%) Specificity (%) FPs/scan
SVM with hand-crafted features 81.5 78.2 6.8
Faster R-CNN (baseline) 88.3 84.6 5.2
3D U-Net 90.1 87.3 4.5
LungNet (proposed) 94.2 91.7 4.0
3.3 Radiologist Study Results
Mean reading time per case: unaided 8.4 min (SD 2.2), with LungNet 5.1
min (SD 1.5) → 39% reduction (p < 0.001, paired t-test).
Sensitivity of radiologists: unaided 86.2%, with assistance 91.5% (p = 0.012).
Inter-observer agreement (Fleiss’ kappa): unaided 0.72 (moderate), with
assistance 0.85 (strong).
Radiologists rated the heatmaps as “very helpful” (4.6/5) for nodule
localisation.
4. Discussion
Our LungNet model demonstrates that a well-designed deep learning system
can substantially improve lung nodule detection and malignancy assessment.
The key strengths are: (a) the attention mechanism that focuses on nodule-like
patterns, (b) the feature pyramid that captures multi-scale information, and (c)
the explainable heatmaps that build clinician trust. The radiologist study is
particularly important – many papers report technical metrics but ignore
real-world workflow integration. A 39% reduction in reading time could
alleviate radiologist burnout and allow more patients to be screened.
Limitations
The dataset is retrospective and may not represent all CT scanner models or
patient demographics. The malignancy classification relies on
radiologist-derived ground truth rather than histopathology. Future work
should include prospective validation on multi-centre data and integration
with electronic health records.
5. Conclusion
LungNet offers a high-performance, explainable AI solution for early lung
cancer detection from CT scans. When used as a second reader, it reduces
reading time and improves diagnostic accuracy. This has the potential to make
lung cancer screening more effective and accessible.