FSP-DETR: Few-Shot Prototypical Parasitic Ova Detection

TL;DR

FSP-DETR tackles the challenge of biomedical object detection under scarce supervision and the presence of unseen classes by unifying few-shot detection, open-set recognition, and zero-shot generalization within a single DETR-based framework. It uses a class-agnostic DETR backbone to propose ova-like regions and builds class prototypes from a small support set, while augmentations generate query embeddings; an energy-based objective combines prototype matching, KL-based calibration, and alignment-driven separation to learn robust embeddings. The method achieves strong gains in low-shot and open-set scenarios across ova, blood cell, and malaria datasets, and demonstrates cross-domain transfer without retraining. A new ova-detection benchmark with 20 parasite classes supports standardized evaluation, underscoring the framework’s practical impact for data-scarce biomedical diagnostics and surveillance.

Abstract

Object detection in biomedical settings is fundamentally constrained by the scarcity of labeled data and the frequent emergence of novel or rare categories. We present FSP-DETR, a unified detection framework that enables robust few-shot detection, open-set recognition, and generalization to unseen biomedical tasks within a single model. Built upon a class-agnostic DETR backbone, our approach constructs class prototypes from original support images and learns an embedding space using augmented views and a lightweight transformer decoder. Training jointly optimizes a prototype matching loss, an alignment-based separation loss, and a KL divergence regularization to improve discriminative feature learning and calibration under scarce supervision. Unlike prior work that tackles these tasks in isolation, FSP-DETR enables inference-time flexibility to support unseen class recognition, background rejection, and cross-task adaptation without retraining. We also introduce a new ova species detection benchmark with 20 parasite classes and establish standardized evaluation protocols. Extensive experiments across ova, blood cell, and malaria detection tasks demonstrate that FSP-DETR significantly outperforms prior few-shot and prototype-based detectors, especially in low-shot and open-set scenarios.

Figures (3)

• Figure 1: Overview. Our proposed FSP-DETR framework provides a simple, modular unified framework for robust few-shot detection from limited support examples, detection of novel/unseen entities through open-set reasoning, and cross-task transfer to new biomedical domains without extensive tuning.
• Figure 2: Framework Overview. The proposed method uses a class-agnostic DETR backbone to extract object proposals from both original and augmented images. Features from the original images are used to construct class prototypes, while augmented image features are passed through a trainable embedding module and classifier. The training objective jointly optimizes prototype alignment and refinement losses. At inference, the model can recognize seen and novel categories by distance-based classification with learned prototypes.
• Figure 3: Qualitative Visualization of FSP-DETR's performance across three tasks: few-shot, zero-shot, and open-set ova detection. Each column shows a representative example, with the top row showing ground truth and the bottom showing predicted detections.