Modular Deep Active Learning Framework for Image Annotation: A Technical Report for the Ophthalmo-AI Project

TL;DR

This work tackles the high cost of manual annotation in retinal OCT image segmentation by introducing MedDeepCyleAL, an end-to-end modular active learning framework for efficient annotation and DL-based segmentation in ophthalmology. It integrates a four-component architecture (Annotation Tool, Controller, Data Manager, Active Learning Backend) with a VISSL-based segmentation backend and modAL-driven AL, all configurable via YAML. The segmentation backbone employs transfer-learning on the Annotated Retinal OCT Images Database (AROI) with UNet variants, and sampling experiments show that EdgeAL and MAXRPR perform best, achieving near $0.99$ accuracy with as little as $12%$ labeled data. Overall, the approach reduces annotation burden and enables interactive diagnostic decision support for AMD and diabetic retinopathy, with strong potential for real-world clinical deployment.

Abstract

Image annotation is one of the most essential tasks for guaranteeing proper treatment for patients and tracking progress over the course of therapy in the field of medical imaging and disease diagnosis. However, manually annotating a lot of 2D and 3D imaging data can be extremely tedious. Deep Learning (DL) based segmentation algorithms have completely transformed this process and made it possible to automate image segmentation. By accurately segmenting medical images, these algorithms can greatly minimize the time and effort necessary for manual annotation. Additionally, by incorporating Active Learning (AL) methods, these segmentation algorithms can perform far more effectively with a smaller amount of ground truth data. We introduce MedDeepCyleAL, an end-to-end framework implementing the complete AL cycle. It provides researchers with the flexibility to choose the type of deep learning model they wish to employ and includes an annotation tool that supports the classification and segmentation of medical images. The user-friendly interface allows for easy alteration of the AL and DL model settings through a configuration file, requiring no prior programming experience. While MedDeepCyleAL can be applied to any kind of image data, we have specifically applied it to ophthalmology data in this project.

Figures (8)

• Figure 1: Design of the AL cycle and the functionality of the individual components: The first step consists of the annotation of an initial data set through the Annotation Tool (AT). The annotation information is then forwarded to a Controller, which initiates the update of the data set (2.) and triggers the next AL learning cycle (3.). The back-end then trains a deep learning model and performs the querying of the non-annotated data (4.). The top $n$ samples resulting from the querying are then sent to the controller, which in turn sends the information to the AT (5). With the help of a data preparation tool, the AT then downloads new samples (6.) and generates preliminary annotations with a pre-trained model. Those preliminary annotations serve to facilitate the annotation process by proposing labels/segmentation, thus reducing annotation costs.
• Figure 2: OCT: a) SLO with the positioning of the exposure on the retina (white square) and the position of a slice (red line), and b) corresponding slice
• Figure 3: Architecture of the Active Learning Infrastructure
• Figure 4: Visual representation of U-net architecture ronneberger_2015
• Figure 5: Basic architecture of HUMAN: Administrators define a protocol that is translated into a state machine. The State Machine interacts with API, Database and Client GUI. Annotators annotate using the GUI.