Danny Xie

Healthcare systems are currently adapting to digital technologies, producing large quantities of novel data. Based on these data, machine-learning algorithms have been developed to support practitioners in labor-intensive workflows such as diagnosis, prognosis, triage or treatment of disease. However, their translation into medical practice is often hampered by a lack of careful evaluation in different settings. Efforts have started worldwide to establish guidelines for evaluating machine learning for health (ML4H) tools, highlighting the necessity to evaluate models for bias, interpretability, robustness, and possible failure modes. However, testing and adopting these guidelines in practice remains an open challenge. In this work, we target the paper-to-practice gap by applying an ML4H audit framework proposed by the ITU/WHO Focus Group on Artificial Intelligence for Health (FG-AI4H) to three use cases: diagnostic prediction of diabetic retinopathy, diagnostic prediction of Alzheimer’s disease, and cytomorphologic classification for leukemia diagnostics. The assessment comprises dimensions such as bias, interpretability, and robustness. Our results highlight the importance of fine-grained and caseadapted quality assessment, provide support for incorporating proposed quality assessment considerations of ML4H during the entire development life cycle, and suggest improvements for future ML4H reference evaluation frameworks.

This research introduces a simple framework for Object Detection (OD) based on few-shot methods and Visual Foundation Models (VFM). The framework comprises of three core modules: (i) object proposal, (ii) embedding creation, and (iii) object classification. We evaluated six distinct VFMs to generate the object proposals. We compared the performances of four feature extractors to optimize the object representation, including convolutional neural networks and transformer-based models. Furthermore, we investigated four few-shot methods for classifying objects using minimal labeled data. Our framework provides a scalable and cost-effective solution, specifically applied to OD for pineapple localization in the drone imagery of large pineapple fields, where labeled data are scarce and expensive.

We introduce PineSORT, a novel Multiple Object Tracking (MOT) system for drone-based agricultural monitoring, specifically tracking pineapples for yield estimation. Our approach tackles key challenges such as repetitive patterns, similar object appearances, low frame rates, and drone motion effects. PineSORT enhances the tracking accuracy with motion direction cost, camera motion compensation, a three-stage association strategy, and overlap management. To handle large displacements, we propose an ORB-based camera compensation technique that significantly improves the Association Accuracy (AssA). Evaluated via 5-fold cross-validation against BoTSORT and AgriSORT, PineSORT achieves statistically significant gains in our Identity-Switch Penalized IDF1 (ISP-IDF1) metric, along with gains in IDF1 (Identity F1 Score), HOTA (Higher Order Tracking Accuracy) and AssA. These results confirm its effectiveness in tracking low-FPS drone footage, making it a valuable tool for precision agriculture.