TLDR: This research introduces the first application of Continual Learning (CL) to Visual Anomaly Detection (VAD) in the medical domain. It addresses the challenge of models forgetting previous knowledge when adapting to new medical imaging data. By utilizing PatchCoreCL, a CL variant of the PatchCore model, and evaluating it on the BMAD dataset, the study demonstrates that this approach effectively identifies anomalies with performance comparable to traditional task-specific models, while significantly reducing forgetting and managing memory efficiently. This paves the way for more adaptive and robust AI in medical diagnostics.
In the rapidly evolving field of medical imaging, the ability to accurately detect anomalies is paramount. Visual Anomaly Detection (VAD) systems are designed to identify unusual patterns in images, relying solely on examples of normal data during their training phase. This approach has proven invaluable in various sectors, including manufacturing, and is increasingly critical in medicine where precise and explainable detection can significantly impact patient outcomes.
However, a significant challenge arises when the characteristics of input data change over time. Traditional VAD models, like many machine learning systems, are susceptible to what’s known as ‘Catastrophic Forgetting.’ This means that when a model learns a new task or adapts to a new type of data, it tends to forget the knowledge it acquired from previous tasks. Given the dynamic nature of medical imaging data – new equipment, different patient populations, or evolving disease presentations – this forgetting can severely degrade a model’s performance and reliability.
This is where Continual Learning (CL) comes into play. CL is a branch of machine learning that enables models to learn new information incrementally without losing previously acquired knowledge. It allows AI systems to adapt to new tasks or data domains over time, avoiding the need for complete retraining on all old and new data simultaneously, which can be computationally intensive and often impractical.
A groundbreaking study, titled Towards Continual Visual Anomaly Detection in the Medical Domain, explores for the first time the application of VAD models within a Continual Learning framework specifically for the medical field. The research, conducted by Manuel Barusco, Francesco Borsatti, Nicola Beda, Davide Dalle Pezze, and Gian Antonio Susto from the University of Padova, addresses this critical gap.
The Approach: PatchCoreCL and BMAD Dataset
The researchers utilized a Continual Learning version of a well-established VAD model called PatchCore, which they refer to as PatchCoreCL. PatchCore is known for its effectiveness in industrial anomaly detection and its ability to provide both image-level anomaly scores and interpretable heatmaps that pinpoint anomalous regions at a pixel level.
To adapt PatchCore for continual learning, several modifications were made. The core idea behind PatchCoreCL is to maintain separate ‘memory banks’ for each task encountered during training. To prevent unbounded memory growth, the total number of stored patch feature vectors is kept fixed. This means that as new tasks are learned, the memory allocated to each individual task is adjusted using a technique called coreset subsampling, ensuring that the most relevant information from past tasks is retained.
The performance of PatchCoreCL was evaluated using BMAD (Benchmarks for Medical Anomaly Detection), a real-world medical imaging dataset. BMAD is a diverse collection of medical imaging datasets spanning multiple anatomical regions and image acquisition modalities, including categories like “Brain AD,” “Liver AD,” “Retina RESC AD,” “Chest AD,” “Histopathology AD,” and “Retina OCT2017 AD.” The dataset was structured into a sequence of tasks to simulate a continual learning scenario, where the model incrementally learns to detect anomalies across these different medical categories.
Also Read:
- BRAIN: Learning from Evolving Brain Signals for Vision Understanding
- Unifying Visual Perception: A Deep Dive into Open World Detection
Key Findings and Impact
The results of the study are highly promising. PatchCoreCL demonstrated performance comparable to ‘task-specific’ models – systems trained independently for each medical category, which typically represent the upper bound of accuracy but are impractical for dynamic environments due to high memory consumption. Crucially, PatchCoreCL achieved this with a ‘forgetting value’ of less than 1%, indicating its remarkable ability to retain previously acquired knowledge while learning new tasks.
The study compared PatchCoreCL with different memory configurations (PatchCoreCL-10k and PatchCoreCL-30k, referring to the memory bank size) against other strategies like ‘Multi-Model’ (separate models for each task), ‘Joint-Train’ (a single model trained on all data at once), and ‘Fine-Tuning’ (naive sequential updates without forgetting mitigation). PatchCoreCL-30k replicated the performance of the Joint-Train approach with efficient memory usage, while PatchCoreCL-10k, using even less memory, still showed very strong results with a minimal performance gap and low forgetting.
This research highlights the feasibility and significant potential of Continual Learning for adaptive Visual Anomaly Detection in medical imaging. By providing a robust baseline for future research, this work paves the way for developing more resilient and adaptable AI models that can continuously learn from new medical data without suffering from catastrophic forgetting, ultimately leading to more accurate and timely diagnoses in clinical settings.
