Respira-Map

Respira-Map

What inspired us: COPD is usually found too late. By the time a breathing test shows a problem, the lungs are already damaged and nothing can undo it. The damage starts years before.

Doctors can already see early damage on a CT scan. But today that gets written down as one word: emphysema, yes or no. Two patients can both be a "yes" and have very different lungs. That one word throws away everything in between. We wanted to replace that word with a number.

What we learned: A lot of it was about the data, not the model. The clinical file uses commas for decimals, except one column that uses a dot. Read it the normal way and the numbers quietly turn into text. Parametric response map (PRM) of patient 17 was identified quite late and we lost time to that before we caught it. We also found that the column marking emphysema is part of how the patient groups are defined. If you use it to predict the group, you get a perfect score that means nothing. We added a check that refuses to let it into the model. The biggest lesson: there are only 14 CT scans. We had built a deep learning model expecting far more. We put it aside. A fancy model trained on 14 scans gives a number nobody should trust. How we built it: We measure the CT scan automatically: how much of the lung is low density, how thick the airway walls are, and dysanapsis — how big the airways are compared to the lung. That last one is strongly linked to COPD risk and most people don't measure it.

We also use the clinical data, with follow-up visits years apart, to see who is losing lung function fastest. Then we check something simple: does the CT tell us anything the clinic doesn't already know from age, smoking and breathing tests? We fit the clinical model first and ask what the scan adds on top. It runs as one pipeline on the supercomputer, and the results come to a laptop so the demo works with no internet. -- Accomplishments that we're proud of We could build fully automated pipeline for training large image corpus. -- What's next for Respira-Map Implementation for Multimodal dataset