Cardiac MRI Classification

This project sits in that slightly strange space between academic ML work and engineering discipline.

The question itself was simple enough: given a reasonably large Cardiac MRI image dataset, how do different deep learning architectures behave when they are pointed at the same classification problem?

The actual work, as usual, was not that simple.

The Setup

The dataset had roughly 63,000 Cardiac MRI images, and I compared multiple transfer-learning architectures across the same broad task framing:

• MobileNetV2
• ResNet152V2
• DenseNet201
• InceptionV3

The idea was not to worship one architecture as the winner before training even started. It was to make the comparison visible: how each model trained, where it struggled, what changed with augmentation and weighting, and whether attention-based additions improved the story or just made the pipeline look fancier.

That last bit matters because ML projects can get performative very quickly. Add a block, add a graph, add a term from a paper, and suddenly it looks more sophisticated without necessarily becoming better.

I wanted the work to be inspectable, not just impressive-looking.

The Training Story

The project explored transfer learning, data augmentation, class weighting, a spatial attention block, and an attention ensemble. The useful part was seeing how those choices changed the behavior of the models rather than treating training as a black box.

Medical-image classification is also a good reminder that accuracy is never the whole story. The evaluation has to be looked at carefully because false confidence is especially dangerous in health-related domains, even in a research or comparative-study setting.

So the project includes saved checkpoints, evaluation plots, reproducible training flows, and the accompanying paper/documentation so the results can be examined rather than simply asserted.

What I Took From It

This was one of the projects that made me more careful about applied ML.

It is easy to get a notebook running. It is harder to build a training and evaluation pipeline that another person can reason about. The gap between those two things is where a lot of real ML engineering lives.

The interesting part in this one wasn't just comparing models but was learning how much structure you need around the comparison before the result deserves to be trusted.

That lesson shows up in a lot of my later work: less blind model worship, more attention to preprocessing, evidence, repeatability, and whether the whole pipeline actually says what we think it says.