Improving necrosis estimation in MR-guided prostate focused ultrasound ablations with a probabilistic thermal dose model

False positive necrosis classification close to ablation zones in magnetic resonance (MR)-guided transurethral ultrasound ablation of the prostate poses an issue for real-time assessment of the treatment. Typically, apriori knowledge about transducers is utilized to filter temperature maps, thereby reducing false positives. Our innovation, however, lies in the ability of our probabilistic CEM 43 thermal dose model (pCEM 43 ) to decrease false positives without the need for prior knowledge beyond the noise statistics estimated from baseline images. The original pCEM 43 was modified improving statistical thermal dose estimation and applied to transurethral ultrasound ablations. The modified model was evaluated on 22 ablation data sets to determine its accuracy and computation time. MR thermometry was acquired by the TULSA Pro system with an EPI sequence using the proton resonance frequency shift (PRFS). Accuracy was determined with Sørensen–Dice coefficients (DSC), relative false positive rates (rFPR), sensitivities. Additionally, over- and underestimation of ablation volumes was investigated with mean contour overestimation (MCO) and mean contour underestimation (MCU) metrics. Results of pCEM 43 were compared to an unfiltered CEM 43 (CEM 43 ) and a spatiotemporally filtered CEM 43 model (STF-CEM 43 ). Computation times were measured for each data set and compared between models. Median DSCs were 38.9% (32.4%–49.3%), 64.2% (53.7%–73.6%), and 70.7% (66.6%–76.3%) ( p ≤ 0 . 001 ) for the CEM 43 , the STF-CEM 43 , and the pCEM 43 model, respectively. Mean rFPRs were 2.18 ±1.03, 0.61 ±0.39, and 0.32 ±0.28 ( p ≤ 0 . 001 ), and median sensitivities were 75.6% (71.7%–83.1%), 72.0% (66.0%–76.2%), and 70.6% (65.1%–74.5%) ( p ≤ 0 . 001 ), respectively. Median MCOs were 21.3 (13.7 - 43.0), 1.1 (0.6 - 2.1), and 0.5 (0.3 - 0.8) ( p ≤ 0 . 001 ) for the CEM 43 , the STF-CEM 43 , and the pCEM 43 model, respectively. Mean MCUs were 0.6 ±0.4, 0.5 ±0.2, and 0.5 ±0.2 ( p = 0 . 025 ). The modified pCEM 43 improves monitoring accuracy and reduces overestimation of ablation volumes without the need for hardware-specific apriori knowledge in real-time. This could increase success rates of treatments and reduce the risk of underablation and subsequent recurrence of tumors. To determine the generalizability of the proposed model, future work will be concerned with performing evaluations for different thermoablation techniques in different types of tissue.