Efficient Precision Control in Object Detection Models for Enhanced and Reliable Ovarian Follicle Counting

TL;DR

The paper tackles the challenge of reproducibly counting ovarian follicles from high-resolution histology by balancing high recall with a guaranteed, controlled precision in object detection. It introduces a model-agnostic post-processing framework based on Learning Then Test (LTT) and a discrete multiple-testing scheme to ensure $\mathbb{E}[\text{Precision}] \ge P_0$ while maximizing recall, and extends this to a multi-parameter detector $(\lambda,\mu)$ that can incorporate biological context. Key contributions include (i) a two-parameter depth-aware detector that emphasizes biologically plausible peripheral follicle localization, (ii) a learned false-detection classifier to further refine predictions, and (iii) an open dataset and extensive cross-split validation showing improved F1-scores and robust precision guarantees over naive thresholds and single-parameter LTT methods. The approach is model-agnostic and enhances reproducibility and efficiency in follicle counting, enabling broader adoption in fertility research without retraining base models. This framework advances trustworthy AI in histopathology by providing probabilistic guarantees on detection quality and practical gains in counting reliability.

Abstract

Image analysis is a key tool for describing the detailed mechanisms of folliculogenesis, such as evaluating the quantity of mouse Primordial ovarian Follicles (PMF) in the ovarian reserve. The development of high-resolution virtual slide scanners offers the possibility of quantifying, robustifying and accelerating the histopathological procedure. A major challenge for machine learning is to control the precision of predictions while enabling a high recall, in order to provide reproducibility. We use a multiple testing procedure that gives an overperforming way to solve the standard Precision-Recall trade-off that gives probabilistic guarantees on the precision. In addition, we significantly improve the overall performance of the models (increase of F1-score) by selecting the decision threshold using contextual biological information or using an auxiliary model. As it is model-agnostic, this contextual selection procedure paves the way to the development of a strategy that can improve the performance of any model without the need of retraining it.

Figures (3)

• Figure 1: Example of an ovary cut with a zoom on an annotated area with PMF (red), Primary (green) and Secondary (blue) follicles
• Figure 2: Computation of the depth of a box. The box is predicted by the OD model and the contour of the ovary is computed. The contour is then dilated until the box is inside. The depth of the box is then computed as the ratio of the area of the dilated contour over the area of the ovary. Each line represents the contour-line of the dilatation associate to a bounding-box and each box represents a prediction of the model.
• Figure 3: Precision, Recall and F1-score for target precision $P_0 = 0.4$. Left: EfficientDet model. Right: Yolo model. In blue: decision with objectness and depth; in orange: decision with objectness and classification score, in green: LTT decision with objectness only; in red: the naive decision.