Deep Learning Approaches for Morphological Classification of Intestinal Organoids

Organoids, derived from primary donor or stem cells, closely replicate the composition and function of their <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">in vivo</i> counterparts. This quality makes them a reliable model for validating hypotheses on disease-related biological processes and mechanisms. To date, the classification of organoids is performed manually by microscope and, therefore, in a data-driven application, is time-consuming, inaccurate, and difficult to morphological analysis process. The use of deep learning (DL) in organoid image analysis becomes crucial to handle complexity, variability, and large amounts of data efficiently and accurately, overcoming the limitations of traditional image processing approaches. In this paper, five CNN-based DL models such as MobileNet, DenseNet, ResNet, Inception, VGG, and the very recent Vision Transformers (ViT) were analyzed using a publicly available dataset for the morphological classification of intestinal organoids. Additionally, traditional ML models, such as SVM and RF, were tested for comparison using a feature set similar to conventional image processing tools. The systematic performance evaluation is designed to guide users in choosing the most suitable model for processing organoid images. Among all models, ViT achieved the highest accuracy of 86.95%, demonstrating its effectiveness in organoid classification. Inception and DenseNet also exhibited strong performance, with accuracy values of 86.10% and 86.47%, respectively. Rather, SVM and RF performed significantly worse, showing an accuracy approximately 20% lower than the selected DL models. Considering efficiency, ViT had the highest accuracy but required more resources (0.0437s/image, 343 MB), while MobileNet, the lightest model (35.6 MB), had the fastest inference time (0.0063s/image). The findings highlight the potential of DL models in enhancing the accuracy of organoid classification while emphasizing the importance of balancing performance with computational efficiency for real-time applications.