Machine-Learning Assisted Segmentation to Assess the Outcomes of Tissue Ablation with Nanosecond Pulses

High-intensity electrical pulsing has emerged as a robust, versatile, and minimally invasive method for ablating tumors and unwanted tissues. The pulse duration varies from nanoseconds to milliseconds, encompassing three primary ablation pulses: Irreversible electroporation (IRE, us-ms), high-frequency IRE (HFIRE, bipolar $\mu{\mathrm{s}}$), and nanosecond pulses (nsEP). In this study, we concentrated on utilizing a plant tissue model (Yukon Gold) to explore the outcomes of ablations using three different pulse waveforms: unipolar 100 $\mu{\mathrm{s}}$ (IRE), Bipolar ±300 ns, and Unipolar 300 ns. The images of the treated potatoes underwent analysis using a U-net machine learning model trained with manually segmented ablation areas as the ground truth. The treated tissues were evaluated at multiple time points. The trained model’s performance was validated on unseen images, achieving an intersection over union (IoU) of 99.41%. Subsequently, we employed the trained model on unsegmented potato images to identify ablation areas. This study aims to investigate the impact of pulse waveforms on the tumor ablation area’s (AA) size, the homogeneity (HG) of the ablation, and the presence of a distinct boundary, known as boundary gradient (BG). Our findings indicated that the bipolar nanosecond pulse seems to be the optimal waveform for tissue ablation. Additionally, we demonstrated that while the Segment Anything Model (SAM), a large segmentation model developed by Meta, excelled in various zero-shot segmentation applications, it failed to precisely segment the ablation area. This suggests that further efforts are necessary to adapt SAM for specialized tasks.