Optimizing Scanning Acoustic Tomography Image Segmentation With Segment Anything Model for Semiconductor Devices

In recent decades, Scanning Acoustic Tomography (SAT) has become a vital technique for characterizing semiconductor devices in non-destructive evaluation. Precise and efficient segmentation of SAT images is crucial for detecting defects and assessing material properties in the semiconductor industry. However, current manual methods are often expensive and susceptible to human error. This study enhances the segmentation process of SAT images using the deep learning model SemiSA, which is fine-tuned from the Segment Anything model. In our experiments, SemiSA was trained and evaluated on a large-scale dataset from the Ohlabs TSAM-400 system, encompassing various semiconductor devices such as Flip Chip, Power Semiconductor, 6-inch and 12-inch Wafer, Transistor, and Multilayer Ceramic Capacitor. The results demonstrate that SemiSA significantly improves segmentation tasks across all types of SAT images of semiconductor devices. On average, there was a 17.89% enhancement in Dice Similarity Coefficient scores and a 24.26% improvement in Intersection over Union scores across all tasks. Additionally, this work also proposes an efficient framework tailored specifically for SAT images. The main objective of developing this segmentation tool is to provide researchers and experts with a valuable tool for advancing the semiconductor evaluation and quality control field. The code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/ThuHa96/SemiSA</uri>.