An open API service indexing awesome lists of open source software.

https://github.com/apartsinprojects/cxr-selective-rejection

Selective multipathology chest X-ray classification via rejection — revised R3 manuscript + point-by-point response (highlighted edits)
https://github.com/apartsinprojects/cxr-selective-rejection

Last synced: 11 days ago
JSON representation

Selective multipathology chest X-ray classification via rejection — revised R3 manuscript + point-by-point response (highlighted edits)

Awesome Lists containing this project

README

          

# Selective Multipathology Chest X-Ray Classification via Rejection Mechanisms (R3)

**Live paper (GitHub Pages): **

Revised manuscript and reviewer response for the Round-3 revision.

- **[index.html](https://apartsinprojects.github.io/cxr-selective-rejection/)**: the revised manuscript. Tracked-edit highlighting versus the Round-2 version uses green for newly added content and yellow for modified content (prose, captions, and table cells). It carries 7 figures and 4 tables, evaluated on three public datasets (NIH ChestX-ray14 n=11,212, MIMIC-CXR n=8,185, PadChest n=5,000) with bootstrap confidence intervals.
- **[response.html](https://apartsinprojects.github.io/cxr-selective-rejection/response.html)**: the point-by-point response to Reviewer 2's three concerns, each with a "Where in the manuscript" pointer to the relevant section, table, figure, or equation, and a Harvard-style reference list.
- **[manuscript.docx](https://apartsinprojects.github.io/cxr-selective-rejection/manuscript.docx)** / **[manuscript_2col.docx](https://apartsinprojects.github.io/cxr-selective-rejection/manuscript_2col.docx)**: single-column and two-column Word versions with the same green/yellow tracked-edit highlighting and native (OMML) equations, suitable for track-changes review.

Headline: at a calibrated operating point, per-pathology entropy-based abstention improves balanced accuracy for all four target pathologies on NIH (gains 0.04 to 0.09 at 0.70 coverage, 95% CIs excluding zero), and the benefit replicates on MIMIC-CXR, on PadChest, on an independently trained model (including cross-dataset on MIMIC), and under leave-one-dataset-out domain shift.

Research code and reproduction pipeline: .