Abstract
The necessity of large amounts of labeled data to train deep models, especially in medical imaging, creates an implementation bottleneck in resource-constrained settings. In Insite (labelINg medical imageS usIng submodular funcTions and sEmi-supervised data programming), we apply informed subset selection to identify a small number of most representative or diverse images from a huge pool of unlabelled data subsequently annotated by a domain expert. The newly annotated images are then used as exemplars to develop several data programming-driven labeling functions. These labeling functions output a predicted-label and a similarity score when given an unlabelled image as an input. A consensus is brought amongst the outputs of these labeling functions by using a label aggregator function to assign the final predicted label to each unlabelled data point. We demonstrate that informed subset selection followed by semi-supervised data programming methods using these images as exemplars perform better than other state-of-the-art semi-supervised methods. Further, for the first time, we demonstrate that this can be achieved through a small set of images used as exemplars.
Introduction
Deep learning has revolutionized medical imaging, but its reliance on vast quantities of labeled data remains a significant hurdle. In clinical practice, "ground truth" labels require the time and expertise of highly trained specialists, making large-scale annotation prohibitively expensive and slow, particularly in resource-limited environments. While semi-supervised learning (SSL) offers a path forward by leveraging unlabeled data, its performance often degrades when the initial labeled set is extremely small or contains noise. The INSITE framework addresses this "expert bottleneck" by maximizing the utility of a minimal labeling budget.
The INSITE Framework
The INSITE (labelINg medical imageS usIng submodular funcTions and sEmi-supervised data programming) framework operates through a three-stage pipeline designed to optimize the labeling process from a cold start.
- Informed Subset Selection: Rather than selecting images at random, INSITE employs submodular functions to identify the most informative subset for expert annotation. Specifically, Facility Location functions are used to select representative images that "cover" the dataset's distribution, while Log Determinant functions ensure diversity among the selected examples.
- Exemplar-based Labeling Functions (LFs): The few images labeled by experts serve as "exemplars." For any unlabeled image, the system calculates its similarity to these exemplars in a feature space defined by a pre-trained ResNet. These similarities act as soft labeling rules; for instance, an image highly similar to a "diseased" exemplar is likely to be diseased itself.
- CAGE (Generative Model) Aggregation: To resolve conflicts and noise among various exemplar-based suggestions, INSITE utilizes the CAGE framework. CAGE is a generative model that aggregates these "weak" labeling functions by considering both the predicted labels and the underlying similarity scores, producing a high-confidence consensus for the entire dataset.
Experimental Evaluation
The framework was validated on three critical medical imaging benchmarks: the APTOS 2019 dataset (retinal fundoscopy for Diabetic Retinopathy), pediatric Chest X-Ray images (pneumonia diagnosis), and Tuberculosis (TB) X-Ray images. In each case, the task was framed as a binary classification problem to evaluate the system's potential as a screening tool.
The results highlight INSITE's exceptional performance in ultra-low-budget scenarios. On the APTOS dataset, with a budget of only 10 labeled images, the LogDeterminant + CAGE approach achieved an accuracy of 0.89, significantly outperforming random selection combined with CAGE (0.64). On the Chest X-Ray dataset, the FacilityLocation + CAGE approach similarly outperformed several state-of-the-art semi-supervised methods at budgets of 10 and 20 images.
Conclusion
INSITE demonstrates that the selection of labeled data is as crucial as the quantity. By treating individual images as visual heuristics, the framework bridges the gap between human-centric rules and automated computer vision. While traditional SSL methods begin to catch up as the labeling budget increases to 40 images, INSITE remains superior for the "cold-start" problem where expert time is at an absolute premium. Future work will focus on extending the framework to multi-class disease grading and incorporating multi-modal clinical metadata into the labeling functions.
Summary updated