jacobeturpin / cxr14

A study of former and current state of the art image classification methods on the NIH ChestXRay14 Dataset

Hello there! 👋

I'm Jacob, and I like to work with data. I build various projects in the AI products and analytics spaces -- both professionally and in my spare time.

My Day Job

I currently work at Avodah as the head of engineering for the AvodahConnect business unit. I have the pleasure of leading a global engineering team in building AI products that provides solutions for real-time translation, transcription, language preservation, and language education.

What I'm Learning

In the evenings, I'm trying to continuously improve my full-stack engineering skillset to better assist the full-stack teams teams I lead. (Academically and professionally I've spent more time in machine learning than application development.) I'm currently tinkering with small scale applications on Next.js.

What I'm Working On

Outside work and studies, here's what you'll find me doing:

• Tinkering with the latest fad in the machine learning community (right now, it's the open source LLMs and RAG-based patterns)
• Creating toy models for sports analytics
• Working on small contributions to OSS libraries

cxr14's Issues

Remove "No Finding" as a class label

The implementation, as it currently stands, has 15 output labels for the multilabel classification. One of these is the label "No Finding" as presented in the NIH ChestX-Ray14 dataset. It's proposed that the label "No Finding" is removed as a potential binary class. The reasoning is that "No Finding" isn't the presence of a given object/pathology like the other classes. Rather, it's simply the absence of a positive class in any of the other possible labels.