A High-resolution dataset for AI-driven segmentation and analysis of drug-treated breast tumor spheroids.

[BACKGROUND AND OBJECTIVE] Three-dimensional (3D) tumor spheroids are widely adopted in preclinical drug screening for their ability to mimic the complexity of in vivo tumor microenvironments. Nevertheless, the high-throughput analysis of such models, especially for quantifying drug responses, remains a significant challenge. This study aims to introduce a high-resolution, publicly available dataset to facilitate AI-driven segmentation and analysis of drug-treated breast tumor spheroids.

[METHODS] Heterotypic spheroids consisting of MDA-MB-231 breast cancer cells and human fibroblasts were cultured in a microfluidic chip and subjected to either treatment with liposomal doxorubicin or left untreated. Microscopic imaging was conducted over eight consecutive days, resulting in 95 high-resolution images. These were preprocessed and divided into 2980 image tiles (512 × 512 pixels), followed by semi-automated annotation. The dataset was evaluated using three deep learning segmentation models: U-Net, Fully Convolutional Network (FCN), Mask R-CNN, YOLOv12-Seg, and DeepLab. Morphological features extracted from the segmented spheroids were analyzed using both statistical and machine learning techniques.

[RESULTS] Among the models tested, DeepLab achieved the highest segmentation accuracy with a Jaccard index of 91.17 %. Key morphological descriptors-area, perimeter, inradius, and boundary complexity-were extracted and analyzed using Generalized Estimating Equations, revealing statistically significant differences (p < 0.05) between control and treated spheroids. Classification using a Support Vector Machine trained on features reduced via Principal Component Analysis resulted in 96 % accuracy in distinguishing the two groups.

[CONCLUSIONS] The HTS-Seg dataset provides a high-quality image resource with corresponding annotations and morphological features, supporting the development and validation of segmentation and classification models in biomedical image analysis. This work enables more accurate in vitro evaluation of drug effects on 3D tumor spheroid models and contributes to advancements in AI-assisted cancer research.