IDCC-SAM: Automated cell counting in immunocytochemistry using the segment anything model

Abstract

Cell counting in Immunocytochemistry is crucial for biomedical research, aiding the diagnosis and treatment of several diseases such as neurological disorders, autoimmune disorders, and cancer. However, traditional counting methods are manual, time-consuming, and prone to human error. While deep learning methods have improved this problem, they rely on labeled datasets, which are also expensive to obtain, posing challenges for scalability and efficiency.

In this work, we present the Immunocytochemistry Dataset Cell Counting using the Segment Anything Model (IDCC-SAM), a pipeline that leverages a zero-shot approach to segment and count cells in immunocytochemistry cellular images using SAM under unprompted settings with no required manual labels. IDCC-SAM utilizes the Meta AI’s Segment Anything Model (SAM), pre-trained on 11 million images and 1.1billion masks, to segment and count cells in Immunocytochemistry cellular images. By employing a zero-shot approach, IDCC-SAM eliminates the need for manual annotations, enhancing scalability and reducing menial human labeling effort.

Our experimental results demonstrate that IDCC-SAM achieves a new benchmark performance on the IDCIA immunocytochemistry dataset, outperforming the traditional supervised state-of-the-art Mask RCNN model by 9% and UNet by 1%, despite the latter being fine-tuned on labeled examples. Our work shows evidence that SAM has the potential to improve performance on cell counting tasks while reducing the reliance on specialized models and manual annotation efforts, via a zero-shot approach.