Focused Active Learning for Histopathological Image Classification

TL;DR

This paper tackles the challenge of data-efficient histopathological image classification in the presence of artifacts, ambiguities, and severe class imbalance. It introduces Focused Active Learning (FocAL), a probabilistic framework that combines a Bayesian Neural Network with an OoD detector to estimate epistemic, aleatoric, and OoD uncertainties, and uses a specially designed acquisition function to prioritize informative, in-distribution samples. The method is validated on MNIST with artificial perturbations and on the Panda prostate cancer dataset, showing superior data efficiency and better class representation, notably achieving a Cohen's kappa of 0.764 with only ~0.69% of labeled patches. The work demonstrates that disentangling different uncertainty sources and filtering artifacts substantially improves active learning in real-world medical imaging tasks and offers a practical approach for reducing labeling costs in pathology.

Abstract

Active Learning (AL) has the potential to solve a major problem of digital pathology: the efficient acquisition of labeled data for machine learning algorithms. However, existing AL methods often struggle in realistic settings with artifacts, ambiguities, and class imbalances, as commonly seen in the medical field. The lack of precise uncertainty estimations leads to the acquisition of images with a low informative value. To address these challenges, we propose Focused Active Learning (FocAL), which combines a Bayesian Neural Network with Out-of-Distribution detection to estimate different uncertainties for the acquisition function. Specifically, the weighted epistemic uncertainty accounts for the class imbalance, aleatoric uncertainty for ambiguous images, and an OoD score for artifacts. We perform extensive experiments to validate our method on MNIST and the real-world Panda dataset for the classification of prostate cancer. The results confirm that other AL methods are 'distracted' by ambiguities and artifacts which harm the performance. FocAL effectively focuses on the most informative images, avoiding ambiguities and artifacts during acquisition. For both experiments, FocAL outperforms existing AL approaches, reaching a Cohen's kappa of 0.764 with only 0.69% of the labeled Panda data.

Figures (9)

• Figure 1: Model overview of the proposed FocAL method. It consists of three main components: feature extractor, BNN, and OoD detection (blue boxes). The figure shows how the different components are combined for training and acquisition.
• Figure 2: Images with ambiguities and artifacts that should be avoided during AL acquisition. The top row shows MNIST images with the artificial noise types 'Merging', 'Gaussian Blur', and 'Black Dots'. They simulate the artifacts and ambiguities encountered in histopathological images (bottom row) in the Panda dataset. The left Panda patch contains two different classes (Gleason Grade 3 and 4), the middle patch is blurry due to wrong microscope focus and the right patch is covered by pen marker, obscuring most tissue parts. Although a clear categorization is difficult, we propose the following scale: The images on the left side show ambiguities but the images are in-distribution because their appearance (color distribution and shapes) is normal. The images on the right side can be considered OoD because the color distribution and shapes substantially differ from the 'normal' images of interest. The blurry images are in between these two extremes as the color distribution and appearance is slightly OoD and they contain ambiguities due to blurry edges and patterns. We will see that with the proposed FocAL method, the shown images are avoided thanks to the aleatoric uncertainty and OoD score.
• Figure 3: Feature distribution of the 2000 images ($X_\mathit{train}$ and $X_\mathit{pool}$) after the last acquisition step with 200 labeled images. Each point represents the feature vector $z$ of an MNIST image, reduced to two dimensions by t-SNE. The distribution supports our categorization of artifacts and ambiguities (Fig. \ref{['fig:noisy_image_examples']}). The images with 'black dots' (depicted as squares) are OoD while the 'merged' images are ambiguous and therefore close to the class boundaries. Blurred images show both characteristics (OoD and ambiguities) as some images are far away from the distribution of interest, while others lie close to the class boundaries.
• Figure 4: Feature distribution, uncertainty, and acquisition scores for the first acquisition step (best viewed with zoom), similar to Figure \ref{['fig:data_dist']}. Labeled images are dots filled with turquoise (Digit "0"), pink (Digit "1"), or yellow (other digits). Unlabeled images are dots with greyscale color, representing the uncertainty or acquisition score (the higher the brighter). The datapoints with a green edge represent noise-free images (good for training) and datapoints with a red edge represent images with artifacts, blur or ambiguities. For the proposed FocAL method, the epistemic uncertainty (a) measures the image informativeness, but it is easily distracted by artifacts and ambiguities. These uninformative images can be captured by a high aleatoric uncertainty (b) or a high OoD score (c). Therefore, in the final FocAL acquisition (d), 9 noise-free images are acquired (and only 1 ambiguous image). The competing methods EN (e) and BALD (f) in comparison acquire almost only images with artifacts or ambiguities in this step which add less information to the training.
• Figure 5: Results of the MNIST Experiments with mean and standard error of five independent runs. In \ref{['fig:mnist_results_a']} the total acquired noisy images are plotted and in \ref{['fig:mnist_results_b']} and the accuracy. The proposed FocAL algorithm effectively avoids acquiring images with ambiguities and artifacts and shows the strongest performance.