TLDR: SDF-YOLO is a new, lightweight AI framework based on YOLOv11, specifically designed for robust and efficient detection of small, rare mitotic figures in pathology images. It addresses challenges like domain variability through a single detection head, coordinate attention, and improved feature mixing. Tested on human and canine tumor datasets, it achieved competitive accuracy and computational efficiency, showing promise for improving cancer diagnosis by providing a reliable and efficient framework for diverse pathological contexts.
Mitotic figure detection is a critical task in computational pathology, serving as a strong indicator of tumor aggressiveness. Pathologists rely on counting these dividing cells to grade tumors and guide treatment decisions. However, this process is often challenging due to significant variability in tissue samples, which can arise from different scanners, tissue types, and staining protocols. This ‘domain variability’ makes it difficult for automated detection methods to perform consistently. Furthermore, mitotic figures are small and rare targets, and their resemblance to non-mitotic nuclei can lead to inconsistencies in manual counting among pathologists.
To address these challenges, researchers have introduced SDF-YOLO (Single Detect Focused YOLO), a novel and lightweight detection framework. This framework is specifically designed for robustly identifying small, rare targets like mitotic figures, even when faced with diverse domain conditions. SDF-YOLO builds upon the established YOLOv11 architecture but incorporates several task-specific modifications to enhance its performance and efficiency.
Key innovations in SDF-YOLO include a single detection head, which is precisely aligned with the typical scale of mitotic figures. This simplification reduces computational overhead while maintaining sensitivity to these small objects. The framework also integrates Coordinate Attention (CA) to boost positional sensitivity, helping the model accurately pinpoint the location of mitotic figures. Additionally, it features improved cross-channel feature mixing, which stabilizes the box and score regression in its single-scale detection setting.
The effectiveness of SDF-YOLO was rigorously tested on three diverse datasets: MIDOG++, canine cutaneous mast cell tumor (CCMCT), and canine mammary carcinoma (CMC). These datasets encompass both human and canine tumors, introducing a wide range of domain variability, including differences in species, tissue types, and scanner origins. This comprehensive evaluation demonstrates SDF-YOLO’s ability to generalize across different pathological contexts.
In preliminary tests for the MIDOG2025 challenge, SDF-YOLO achieved impressive results. It recorded an average precision (AP) of 0.799, a precision of 0.758, a recall of 0.775, and an F1 score of 0.766. Its detection and localization performance was further highlighted by an FROC-AUC of 5.793. These metrics underscore the framework’s competitive accuracy and computational efficiency, making it a promising tool for real-world pathology applications. While the assessment was limited to the preliminary phase of the challenge, the results indicate that SDF-YOLO provides a reliable and efficient framework for robust mitotic figure detection across diverse domains.
Also Read:
- Adaptive AI Strategies Advance Atypical Mitosis Classification in Pathology
- Advancing Astrocyte Detection with a Hybrid AI Model
For more detailed information, you can refer to the original research paper.
