Funded by: Afeyan Foundation
PI: Aram Butavyan Co-PI: Varduhi Yeghiazaryan
Researcher: Naira Matosyan
The objective of the research is to generate reliable classification outputs for hyperspectral images (HSI) without requiring expert labeling of portions of the data. To address this, we conduct experiments aimed at understanding the extent to which classification can be automated. Specifically, we seek to determine the minimal amount of data required to achieve high-quality classification of previously unseen images across different application domains.
The initial experiments follow a two-step procedure. First, a model is trained on one or several HSI images from a particular domain and then used to generate predictions on new, unseen images. Second, from these predictions, a small subset of pixels with highly confident outputs is selected, and the model is retrained using these pixels as pseudo-labels.
We use the bovine heart dataset (acquired at L.A. Orbeli Institute of Physiology) and perform experiments with the LeViT model, which has shown promising results for HSI tasks. From the prediction outputs, we select pixels with the highest and lowest predicted probabilities, based on the activation function applied to the logits. These two groups are treated as proxies for ablated and healthy tissue samples, respectively. The figure shows two such examples using different proportions of selected pixels (e.g., 1%, 3%, etc.). All experiments are conducted using five different splits, and the reported accuracies are averaged across these splits. We observe that the model is able to identify healthy tissue with nearly 100% accuracy, while
Figure 1. Prediction accuracy on an unseen image for varying proportions of labeled data (ablated vs. healthy) [left], and two example predictions [right]: red—ablated, green—healthy, black—uncertain/background.
performance on ablated regions remains comparatively low.
Overall, the results suggest that this approach is capable of learning useful patterns and generalizing to unseen images. However, additional experiments are needed to obtain more reliable accuracy estimates. As a next step, we plan to explore self-supervised approaches to further improve performance.