Musca: Deep Learning-Based Web Application for Counting Laboratory Populations of Drosophila melanogaster

Godiksen, Christian Schmidt ; Knudsen, Kristoffer Nøddebo

4. term

Computer Science, Master

2022

2022-06-16

31 pages

The fruit fly Drosophila melanogaster has seen extensive use in scientific research due to its application as a versatile model organism. However, conducting large-scale experiments with Drosophila is time-consuming, potentially requiring the manual counting of Drosophila eggs and larvae in thousands of Petri dishes. In recent years, machine learning, particularly deep learning, has been a groundbreaking technology for alleviating manual labor. We explore the state-of-the-art within deep learning for object detection and conduct a comparative study to select the optimal model for this domain. Leveraging the state-of-the-art deep learning model for object detection, YOLOv5m, we devise an innovative web application that will give researchers the ability to automatically detect and classify objects in Drosophila image data. On our 93 image test set, we achieve a mAP@0.50 of 0.724 and mAP@[0.50:0.95] of 0.543 under a confidence threshold of 0.25 and an NMS IOU threshold of 0.45 with an inference time of 0.3 seconds per image. Further, when treating the detector as class-agnostic while ignoring the predicted bounding box and only counting the detections, we achieve a MAE of 5.305. The proposed software solution provides an opportunity to utilize the cutting-edge within deep learning without requiring prior knowledge of the field and tedious setup from the user. Specifically, the software fulfills the base requirements for supporting Drosophila research and provides extended functionality to increase usability. Domain-specific issues, such as diverse data setups and clumping of objects, are handled explicitly. Further, the system provides granular control of users and resources to function optimally as a production software system.

The fruit fly Drosophila melanogaster has seen extensive use in scientific research due to its application as a versatile model organism. However, conducting large-scale experiments with Drosophila is time-consuming, potentially requiring the manual counting of Drosophila eggs and larvae in thousands of Petri dishes. In recent years, machine learning, particularly deep learning, has been a groundbreaking technology for alleviating manual labor. We explore the state-of-the-art within deep learning for object detection and conduct a comparative study to select the optimal model for this domain. Leveraging the state-of-the-art deep learning model for object detection, YOLOv5m, we devise an innovative web application that will give researchers the ability to automatically detect and classify objects in Drosophila image data. On our 93 image test set, we achieve a mAP@0.50 of 0.724 and mAP@[0.50:0.95] of 0.543 under a confidence threshold of 0.25 and an NMS IOU threshold of 0.45 with an inference time of 0.3 seconds per image. Further, when treating the detector as class-agnostic while ignoring the predicted bounding box and only counting the detections, we achieve a MAE of 5.305. The proposed software solution provides an opportunity to utilize the cutting-edge within deep learning without requiring prior knowledge of the field and tedious setup from the user. Specifically, the software fulfills the base requirements for supporting Drosophila research and provides extended functionality to increase usability. Domain-specific issues, such as diverse data setups and clumping of objects, are handled explicitly. Further, the system provides granular control of users and resources to function optimally as a production software system.

YOLOv5 ; Drosophila ; Object detection ; Classification ; Bioscience ; Object counting ; Fruit fly ; R-CNN ; RetinaNet ; EfficientDet

Download
View record in AAU Student Projects