Transfer learning from RF to B-mode temporal enhanced ultrasound features for prostate cancer detection

Abstract

Purpose: We present a method for prostate cancer (PCa) detection using temporal enhanced ultrasound (TeUS) data obtained either from radiofrequency (RF) ultrasound signals or B-mode images.

Methods: For the first time, we demonstrate that by applying domain adaptation and transfer learning methods, a tissue classification model trained on TeUS RF data (source domain) can be deployed for classification using TeUS B-mode data alone (target domain), where both data are obtained on the same ultrasound scanner. This is a critical step for clinical translation of tissue classification techniques that primarily rely on accessing RF data, since this imaging modality is not readily available on all commercial scanners in clinics. Proof of concept is provided for in vivo characterization of PCa using TeUS B-mode data, where different nonlinear processing filters in the pipeline of the RF to B-mode conversion result in a distribution shift between the two domains.

Results: Our in vivo study includes data obtained in MRI-guided targeted procedure for prostate biopsy. We achieve comparable area under the curve using TeUS RF and B-mode data for medium to large cancer tumor sizes in biopsy cores (>4 mm).

Conclusion: Our result suggests that the proposed adaptation technique is successful in reducing the divergence between TeUS RF and B-mode data.

Keywords: B-mode; Cancer diagnosis; Deep belief network; Deep learning; Prostate cancer; Radiofrequency signal; Temporal enhanced ultrasound; Transfer learning.

Fig. 1

Temporal enhanced ultrasound (TeUS) data generation. Changes in the backscattered response from a point in tissue (shown in red dot) over a sequence of US frames are analyzed to differentiate tissue types

Fig. 2

An illustration of the proposed approach for domain adaptation between RF and B-mode time series data in TeUS framework

Fig. 3

Learning curve for DBN training based on the cross-entropy: a for the first hidden layer size. b For different learning rates (LR). c For different mini-batch size (BS). In a coarse search for the meta-parameters, we achieved the lowest cross-entropy loss with n = 44, LR = 0.2, and BS = 10

Fig. 4

Distribution shift from B-mode to RF for the top three features before (top row) and after (bottom row) the shared deep network. The proposed domain adaptation method can effectively align features and reduce the distribution shift in common learned feature space, a distribution of feature 1, b distribution of feature 2, and c distribution of feature 3

Fig. 5

Influence of labeled dataset size in classification accuracy: performance of the method measured by AUC, accuracy, sensitivity, and specificity in the k-fold cross-validation setting for a TeUS RF data and b TeUS B-mode data

Fig. 6

Comparative performance of the proposed method measured by AUC over the baselines for D_test

Fig. 7

Cancer probability maps overlaid on B-mode US image, along the projected needle path in the temporal US data and centered on the target. The ROIs for which the cancer likelihood are more than 50% are colored in red; otherwise, they are colored as blue. Red boundary shows the segmented prostate in MRI projected in TRUS coordinates, dashed line shows needle path, and the arrow pointer shows the target: a—c correctly identified cancerous core using RF time series data; b—d correctly identified cancerous core using B-mode time series data