ANN-Based Computer Aided Diagnosis: From Clinical Second Opinions to Educational Transformation

TL;DR

This research investigates the dual-role of Artificial Neural Networks (ANNs) in the field of medical imaging. By moving away from "hand-crafted" features toward Pixel-Based Machine Learning (PML), the project aims to develop a Computer-Aided Diagnosis (CAD) system that not only assists radiologists in detecting pathologies like lung nodules and breast cancer but also serves as a specialized educational platform for medical trainees.

Background & Motivation

The explosion of high-resolution imaging modalities (CT, MRI) has led to an unprecedented increase in radiologist workload. This "data deluge" often results in diagnostic fatigue and variability. Traditional CAD systems, while helpful, often required manual feature extraction—a bottleneck that limited their flexibility. The author’s insight is that Deep Learning and CNNs can bypass this by learning directly from raw pixel data, effectively acting as an automated "second pair of eyes."

Methodology: The Shift to Pixel-Based Learning

The core of this work lies in the implementation of Deep Convolutional Neural Networks (CNNs) and Massive-Training ANNs (MTANNs).

1. Architectural Insights

Unlike traditional machine learning which requires manual segmentation, these architectures utilize:

• Convolutional Layers: To automatically extract spatial hierarchies and textures.
• Max Pooling: To ensure translational invariance, allowing the model to recognize a lesion regardless of its exact location in the frame.
• Data Augmentation: To overcome the scarcity of medical data, the research utilizes geometric transformations to "expand" small datasets like the JSRT (chest radiographs).

2. Implementation Framework

The project leverages the MatConvNet (Matlab) and Theano (Python) ecosystems. This hybrid approach allows for rapid prototyping and GPU-accelerated training, essential for the high-dimensional nature of CT and MRI volumes.

Note: The CAD design process integrates public database acquisition (LIDC-IDRI, DDSM) with iterative ANN training.

Clinical vs. Educational Validation

A unique aspect of this dissertation is the Second Stage focus:

• Phase 1 (Diagnostic): Evaluating sensitivity, specificity, and False Positive (FP) rates. The goal is to match or exceed human-only performance.
• Phase 2 (Educational): Investigating if trainees who use CAD as a "feedback loop" develop diagnostic skills faster than those using traditional methods. This addresses the "black box" criticism of AI by using the system to highlight error-making patterns in students.

Critical Analysis & Future Outlook

The primary challenge identified in this work is generalization. An ANN trained on one database (e.g., DDSM for mammography) must be robust enough to handle noise from different hospital scanners.

Conclusion

The transition from CAD as a tool to CAD as a teacher marks a significant shift in medical AI. By integrating these systems directly into the Picture Archiving and Communication Systems (PACS), the research paves the way for a future where AI handles the "heavy lifting" of detection, allowing specialists to focus on complex diagnostic decisions and patient care.

Limitations

• High dependency on the quality of expert annotations in public databases (ground truth).
• The "person-dependent" nature of the learning process makes the educational validation phase inherently complex to quantify.

Future work focuses on expanding these methodologies to interstitial lung diseases and brain tumor segmentation.