Leveraging deep learning for accurate lung cancer detection
This research falls in the Artificial Intelligence in Medical Imaging applications to Computer-Aided Diagnosis (CAD) category. Lung cancer is one of the leading global causes of death due to cancer, where survival directly depends on early detection and accurate tumour localisation [2]. Conventional...
Saved in:
| Main Author: | |
|---|---|
| Format: | Final Year Project / Dissertation / Thesis |
| Published: |
2025
|
| Subjects: | |
| Online Access: | http://eprints.utar.edu.my/7007/1/fyp_IA_2025_LLJ.pdf http://eprints.utar.edu.my/7007/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This research falls in the Artificial Intelligence in Medical Imaging applications to Computer-Aided Diagnosis (CAD) category. Lung cancer is one of the leading global causes of death due to cancer, where survival directly depends on early detection and accurate tumour localisation [2]. Conventional pipelines rely heavily on manual interpretation of CT scans, which are time-consuming, subjective, and prone to errors [2]. Modern CAD frameworks also have limitations, with classification-only pipelines lacking localisation, segmentation-first pipelines squandering computation, and parallel pipelines introducing redundancy [15]. To address these problems, this project proposes a classification-first pipeline with EfficientNetB1 for lung cancer classification and conditional Fuzzy C-Means (FCM) segmentation module activated only after cancer detection. The system was developed and deployed via an interactive Streamlit dashboard, which integrates multiple functions like classification, segmentation, lesion measurement, structured reporting (PDF and AIM XML), review by radiologists, and storing cases with SQLite. The workflow included preparation of the dataset from the Kaggle 2D Chest CT Scan dataset, strict preprocessing and augmentation, and comparison-based evaluation of EfficientNetB0 and EfficientNetB1 models with stratified k-fold cross-validation. EfficientNetB1 was employed as the ultimate classifier with accuracy of 94.60% on the test set. The FCM segmentation produced clinically interpretable tumour overlays, and the dashboard successfully integrated error handling, simulation of workflow, and peer review by radiologists. The project concludes with a light, modular, and clinically useful diagnostic system for the assistance of radiologists in the early detection of lung cancer. The future holds quantitative assessment of segmentation and scaling to PACS integration for clinical deployment. |
|---|