Deep Learning & Computer Vision (7053SCN) - Coursework Repository

Project Overview & Problem Definition

This repository contains the continuous codebase, data pipelines, and deep learning architectures developed for the 7053SCN Written Coursework at Coventry University.

The exclusive focus of this project is Medical Image Classification: engineering an end-to-end deep learning pipeline for the binary classification of Pediatric Chest X-Rays to accurately differentiate between Normal anatomy and Pneumonia. This problem carries immense clinical significance; implementing automated, highly accurate screening models serves as a crucial decision-support tool for medical professionals, potentially accelerating diagnostic workflows in high-pressure healthcare environments.

The core objective of this study is to critically evaluate and compare the performance, optimization convergence, and computational efficiency of two distinct modeling strategies applied to this single diagnostic task:

A custom Convolutional Neural Network (CNN) built from scratch.
A transfer learning architecture utilizing a deep, pre-trained feature extractor (MobileNetV2).

Repository Milestones & Development History

To reflect genuine continuous development throughout the academic block, this repository closely follows the mandatory milestone tracking framework:

initial_commit: Setup of the local workspace environment and commitment of the base repository structure.
milestone1_commit: Successful implementation of the target data pipeline, path verification within the cloud workspace, configuration of data shapes, and visualization of initial training samples.
milestone2_commit: Core structural configuration and training implementation of Model 1 (the baseline custom CNN).
milestone3_commit: Integration of Model 2 (MobileNetV2 transfer learning), extraction of comparative training metrics, and compilation of structural parameter summaries.
final commit: Production-ready, fully documented code state containing fully optimized pipelines, clean inline documentation, and final visualization tracking charts prior to submission.

Data Pipeline & Preprocessing

The diagnostic images are managed using high-performance streaming loaders to maintain strict memory efficiency:

Data Allocation: Images are ingested dynamically from the dedicated training (train) and validation (val) subdirectories.
Tensor Transformations: All input visuals are uniformly resized to target dimensions of 150 x 150 pixels across 3 color channels (RGB) and processed in optimized batch sizes of 32.
Normalization: Pixel values are systematically rescaled between 0 and 1 via an integrated rescaling layer (layers.Rescaling(1./255)) to eliminate gradient instability and ensure smooth numerical convergence during backpropagation.

Model Architectures

Model 1: Custom CNN Built From Scratch

An independent, sequential convolutional pipeline designed to automatically extract local spatial features:

Alternate processing blocks utilizing 2D Convolutional layers (layers.Conv2D) with localized 3x3 filters to map structural anomalies in the lung fields.
Max-Pooling operations (layers.MaxPooling2D) to scale down spatial dimensionality and optimize computational overhead.
Rectified Linear Unit (ReLU) activation functions to introduce necessary non-linearity across layer boundaries.
A dense classification head culminating in a single Sigmoid activated neuron, optimized via binary_crossentropy loss to output explicit diagnostic probabilities.

Model 2: Transfer Learning (MobileNetV2)

A transfer learning framework engineered to leverage heavy features optimized on massive external vision datasets:

Base Feature Engine: Google's MobileNetV2 architecture, utilized with pre-trained weights derived from the ImageNet dataset.
Layer Freezing: The entire pre-trained convolutional base is explicitly frozen (base_engine.trainable = False), locking the sophisticated feature extraction layers to prevent weight distortion while forcing the model to lean entirely on Google's generalized visual representations.
Classification Head: A global average pooling layer (layers.GlobalAveragePooling2D) coupled with a lightweight, high-speed dense classification layer, compiled using the Adam gradient descent optimizer.

Computational & Ethical Tracking

Hardware Optimization: Pipeline executions are accelerated using an allocated cloud-based Nvidia T4 GPU environment to effectively handle heavy matrix operations during training epochs.
Parameter Auditing: Model summary utilities are run natively to contrast total parameter counts against active trainable parameters, tracking the exact computational footprint and memory requirements of each model block.
Ethical Considerations: Diagnostic outputs are framed within strict clinical safety parameters. The codebase accounts for dataset representation constraints, the mitigation of diagnostic bias, and the vital necessity of model interpretability before deploying autonomous tools in medical settings.

AI Use & Academic Integrity Declaration

In compliance with institutional criteria for Amber-category submissions, the operational bounds of AI assistance are transparently outlined below:

Tool	Usage Details within this Assignment
Generative AI (Gemini)	Utilized exclusively as an interactive technical assistant to resolve notebook path errors (`FileNotFoundError`), restore persistent variables following GPU environment restarts, and clean up conditional plotting syntax for the comparative loss curves.
Grammarly	Applied to proofread markdown text blocks, refine documentation syntax, and ensure a formal, academic tone appropriate for a technical repository.

All architectural selections, pipeline logic alterations, and training configurations were manually validated, integrated, and executed by me to fulfill the core learning outcomes of the module.

rufaun/7053SCN-Coursework

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Languages