Towards Confident Detection of Prostate Cancer Using High Resolution Micro-ultrasound

Self-Supervised Learning with Limited Labeled Data for Prostate Cancer Detection in High Frequency Ultrasound

“…Rows 1 and 2 are SL models with either random initialization or ImageNet pretraining. Row 3 is the evidential deep learning (EDL) + co-teaching model of [27]. The only significant difference between the SL models is that the EDL+co-teaching model had superior patch-wise performance.…”

“…For a stronger comparison, we consider several supervised baselines: these include random initialization, ImageNet pretraining, and training with vs. without augmentations. We also compare to the model of Gilany et al [27] which uses supervision in conjunction with co-teaching to handle label noise and uncertainty (this model used the same testing set, making the comparison possible).…”

“…When computing the metrics, we ignore cores with cancer involvements of less than 40%. This choice follows conventions used by the previous literature [25]- [27] and is justified as (1) cores with low involvement cores are less likely to contain features representative of cancer and (2) weak labelling means that using a low involvement core will result in more incorrectly labeled patches than correctly labeled ones. Training with these cores is therefore problematic and performance metrics using them are invalid.…”

“…Clinical studies show that micro-ultrasound has a sensitivity comparable to mp-MRI using the qualitative ultrasound based "PRI-MUS" scoring system [21]- [23], and a recent meta-analysis of 13 published studies with 1125 participants concludes that micro-ultrasound guided biopsy has similar PCa detection rates as mp-MRI fusion biopsy [24]. However, analysis of RF data remain relatively unexplored for micro-ultrasound: these are limited to a single study using machine learning with QUS (Rohrbach et al [25]) and two studies (Shao et al [26], and Gilany et al [27]) using deep learning. While promising, there are key challenges to the development of machine-learning based cancer detection models that are clinically useful.…”

“…Javadi et al propose to use multi-instance learning [32] and co-teaching [33] for PCa detection. Uncertainty estimation allows a model to express uncertainty when seeing OOD data rather than making false predictions; such methods have been applied to prostate [27], [34] and breast [35], [36] ultrasound to address heterogenity. Shao et al [26] propose to handle inter-center distribution shift on micro-ultrasound data by training an adversarial auxiliary model to predict data center origin from the hidden features of a cancer detection model, thereby encouraging the detection model to be invariant under distribution shifts.…”