Automatic delineation of tumor spheroids in microscopic images using deep-learning

Ziel/Aim: Tumor spheroid growth assays are used to evaluate the potential of cancer therapies in-vitro. From these, extensive microscopic image series are generated and commonly analyzed via manual threshold-based delineations. Due to morphological changes of the spheroids during treatment, these threshold-based methods require extensive time-consuming manual corrections. In our work we have evaluated to which extend a deep-learning (DL) method allows to automate and optimize such spheroid growth assays with the ultimate goal to avoid manual delineation and corrections.

Methodik/Methods: Spheroids were grown from mouse pheochromocytoma (MPC) cells and treated with X-ray beam irradiation or with a particle emitting radioligand (177Lu-DOTA-TATE). Spheroid growth was monitored over a period of 35 days with two different microscopes. Data consisting of N=38090 images were divided into two independent datasets: one for network training and internal validation using a 5-fold cross-validation scheme (N=21567) and one for final external testing (N=16523). In both groups, the manual delineation served as ground truth (GT). The network was developed using the nnU-Net v2 DL framework. DL-based delineations were compared to manual delineations using the Dice similarity coefficient (DSC). In addition, treatment effects in one specific spheroid trial were calculated as half-maximum spheroid control doses (SCD50) and compared.

Ergebnisse/Results: The median DSC in the main and testing datasets were 0.979 and 0.974, respectively. In the main dataset, only 9% (N=1904) of the DL generated delineations were considered outliers deviating more than 1.5 times the interquartile range from the first quartile of the DSC distribution. Furthermore, only 5% (N=1051) yielded a DSC below 0.85. Similar values were observed in the testing dataset (10% and 4% for outliers and cases with DSC<0.85) with no notable difference between the two microscopes. The SCD50 was comparable between manual (day13: 0.086±0.001, day35: 0.150±0.001) and DL-based delineation (day13: 0.083±0.002, day35: 0.149±0.007).

Schlussfolgerungen/Conclusions: Despite rare cases of incorrect delineation, our network delivers almost perfect delineation predictions if GT data is carefully chosen and reduces the necessary time for spheroid processing from several days to just a few hours.

Publication History

Article published online:
12 March 2025