1324-GRP RetinaSync: A Deadlock-Free Concurrent Pipeline for Diabetic Retinopathy Screening

Presenter Information

Location

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Document Type

Event

Start Date

20-11-2026 2:00 PM

Description

Deep learning can detect diabetic retinopathy (DR) in fundus images, but most studies test models offline and one image at a time. In practice, many clinics share one server with limited GPU memory, so a naive concurrent design can lose images, read half-updated models, or deadlock. This paper presents RetinaSync, a screening pipeline built on classic operating system ideas. Semaphores coordinate clinics and inference workers as producers and consumers. Each worker screens an image with MobileNetV2 and sends only uncertain images to InceptionResNetV2. A manager based on the Banker’s algorithm grants GPU memory only in safe states, and readers-writers locks allow live model updates. The key insight is that a fixed set of models makes the maximum memory claim of every job known in advance, which makes deadlock avoidance practical. We evaluated RetinaSync on 3,662 APTOS 2019 imagesnwith a real GPU and up to 32 bursty clinics. RetinaSync never deadlocked, lost an image, or tore a model read, while a naive design deadlocked in 37 of 40 runs. Under heavy load, it screened 36% to 45% more images per second than reserving the full memory claim up front, and each safety check cost about 0.2% of the time to screen one image. Finally, every prediction matched offline inference, and the cascade reached 95.8% accuracy while sending only 7.5% of the images to the larger model.

This document is currently not available here.

Share

COinS
 
Nov 20th, 2:00 PM

1324-GRP RetinaSync: A Deadlock-Free Concurrent Pipeline for Diabetic Retinopathy Screening

https://campus.kennesaw.edu/colleges-departments/ccse/events/computing-showcase/fa26-cday-program.php

Deep learning can detect diabetic retinopathy (DR) in fundus images, but most studies test models offline and one image at a time. In practice, many clinics share one server with limited GPU memory, so a naive concurrent design can lose images, read half-updated models, or deadlock. This paper presents RetinaSync, a screening pipeline built on classic operating system ideas. Semaphores coordinate clinics and inference workers as producers and consumers. Each worker screens an image with MobileNetV2 and sends only uncertain images to InceptionResNetV2. A manager based on the Banker’s algorithm grants GPU memory only in safe states, and readers-writers locks allow live model updates. The key insight is that a fixed set of models makes the maximum memory claim of every job known in advance, which makes deadlock avoidance practical. We evaluated RetinaSync on 3,662 APTOS 2019 imagesnwith a real GPU and up to 32 bursty clinics. RetinaSync never deadlocked, lost an image, or tore a model read, while a naive design deadlocked in 37 of 40 runs. Under heavy load, it screened 36% to 45% more images per second than reserving the full memory claim up front, and each safety check cost about 0.2% of the time to screen one image. Finally, every prediction matched offline inference, and the cascade reached 95.8% accuracy while sending only 7.5% of the images to the larger model.