Use off-the-shelf image models to predict the "educational value" of chest x-rays and route high educational value studies to residents

Research Year Project Goal:

Explore and document the capabilities of artificial intelligence in radiology

Explore how humans interact with artificial intelligence in a deployed scenario

Why?

Use off-the-shelf image models to predict the "educational value" of chest x-rays and route high educational value studies to residents

Explore and document the capabilities of artificial intelligence in radiology

Use off-the-shelf image models to predict the "educational value" of chest x-rays and route high educational value studies to residents

Explore how humans interact with artificial intelligence in a deployed scenario

Stage 1: Retrospective analysis of PMBB data to determine the capabilities of open models

Stage 2: Depending on capabilities of open models, work with the chest department and AI Insights Infra to deploy the routing platform

Stage 3: Create HCI/HAII experiments making use of the platform

Stage 1: Model Retrospective Analysis

How to define "educational value"

- By list of important diagnoses

- By list of rare diagnoses

- By list of diagnoses individualized to that resident that the resident hasn't yet seen

- By diagnosis/finding individualized to the resident that they had missed in the past

Take advantage of our Macro-X system to filter for studies that have "Great Call" "Major Change" "Minor Change" "Notification"

Class	Last Year	All Time	PMBB
All	583,468	6,877,112	~2K
Notify	6,044	86,901	1,693
Major	204	1,409	221
Minor	6,337	28,754	2,548
Great	228	1,918	347

Model

AI Insights Infra

Router

Resident

Extender

$0.493 per hour for a NC4as T4 v3 (16gb VRAM)

$3.505 per hour for a NV36ads A10 V5 (24gb VRAM)

*27B VLMS (Would require 80GB VRAM and are therefore too expensive for deployment -- MedGemma-27B)

By Vineeth Gangaram