This repository is dedicated to the study of model robustness and data sensitivity within the development of Computer-Aided Diagnostic (CAD) systems. Our research focuses on the classification of Chest X-rays (CXR) to detect Tuberculosis (TB)—treating it not merely as a standard pattern recognition task, but as a high-stakes clinical application where precision, interpretability, and trust are paramount.
The core of this project is a comparative study of Transfer Learning using deep convolutional neural networks. We utilize baseline models, such as ResNet50, and pivot toward advanced techniques like LoRA (Low-Rank Adaptation) fine-tuning to handle medical data that is exceptionally sensitive to architectural changes.
In medical AI, a "prediction" alone is insufficient. Therefore, this project explores the intersection of deep feature extraction and Explainable AI (XAI). By generating visual evidence through heatmaps, we justify the model’s clinical predictions, building a transparent framework that clinicians can audit, understand, and ultimately rely on for diagnostic support.
This project contributes to the broader field of Computer Science and Medical AI by documenting the transition from "Black Box" models to "Clinically Transparent" systems.
- The Transition to LoRA (Low-Rank Adaptation): Moving beyond standard fine-tuning, we are investigating the application of LoRA for clinical data. By training low-rank matrices while keeping the base model specialized, we aim to capture micro-detail sensitivity without the catastrophic forgetting or overfitting typically seen in small-sample medical datasets.
- Bridging the Trust Gap with XAI: We implemented Grad-CAM (Gradient-weighted Class Activation Mapping) to generate visual evidence. By producing heatmaps that highlight the specific regions influencing a prediction, we transform the model from a "decision-maker" into a "decision-supporter" for radiologists.
- Studying Baseline Errors: By analyzing where ResNet50 failed, we have gained insights into the error surfaces of deep models in clinical contexts. This research is currently feeding into a custom, monochromatic-sensitive model (In Progress) designed to outperform general-purpose architectures.
- Clinical Image Processing: We study Chest X-ray images and use STA image processing to preserve biological characteristics and transfer them to the model sensitively.
- Transfer Learning: Utilizes the ResNet50V2 (Baseline) model pre-trained on ImageNet weights. We identified that ImageNet characteristics differ significantly from CXR data, leading to the development of a main model trained on micro-detail sensitive, monochromatic datasets (In Progress).
- Dual-Stage Training: Employs a robust strategy of (1) Head Training (frozen base) followed by (2) Fine-Tuning (unfrozen top layers) with a very low learning rate.
- Explainable AI (XAI): Implements Grad-CAM to highlight pathological regions of the lung influencing the model's prediction.
- Performance Metrics: Focuses on clinically relevant metrics such as Area Under the Curve (AUC), Recall (Sensitivity), and Precision.
Every significant research breakthrough begins with the identification of a baseline’s limitations. In the initial phase of this project, we employed ResNet50V2—a gold standard in general computer vision—pre-trained on the ImageNet dataset.
The Findings: While ResNet50 achieved respectable metrics on paper, it failed to meet the rigorous standards required for clinical deployment. We identified a fundamental domain gap:
- Semantic Divergence: ImageNet features are optimized for natural objects (texture, color, and macro-shapes like animals or vehicles).
- Clinical Nuance: Chest X-rays are monochromatic, high-bit-depth images where the diagnostic signal lies in micro-details—subtle opacities, pleural thickening, and architectural distortion of lung parenchyma.
This "productive failure" proved that standard Transfer Learning is often too "coarse" for medical imaging, necessitating architectures that respect high-resolution, monochromatic data.
Understanding the data is the first step of the research. We applied Spatio-Temporal Analysis (STA) and specialized image processing to preserve essential biological characteristics.
- Biological Preservation: We focused on maintaining the integrity of the costophrenic angles and hilar regions, ensuring that preprocessing did not artifactually introduce "pathology" or erase subtle lesions.
- Monochromatic Sensitivity: Unlike standard RGB processing, our pipeline was designed to handle the specific intensity distributions found in clinical CXR hardware.
Kaggle: Chest X-ray Dataset (Montgomery and Shenzhen)
Primary Sources:
- Montgomery County CXR Set: Link to Source
- Shenzhen Hospital CXR Set: Link to Source