DeepTOBIDx: deep learning-enhanced multimodal diagnostic breast imaging

PROJECT SUMMARY / ABSTRACT Imaging modalities routinely used in the diagnostic workup, i.e., mammography, ultrasound, and MRI, can catch breast cancers at an early stage based on structural abnormalities but lack in providing physiological information relevant to the function of tissue that determines tumor malignancy. Currently, the U.S. national benchmark of

positive predictive value for malignancy at biopsy after a BI-RADS 4 or 5 diagnostic assessment (PPV3) is only 30.4%. This means about 7 out of 10 biopsies come back negative for cancer. Prior research has demonstrated that diffusion optical tomography (DOT), as a complementary functional imaging modality to clinical breast

imaging, bears ample potential for differentiating malignant and benign breast lesions to reduce unnecessary biopsies. However, the clinical utility of DOT for breast cancer diagnosis is limited by two factors. First, due to limited contrast recovery, conventional DOT has been primarily validated in patients with large masses, leaving

its ability to characterize smaller lesions often seen in the diagnostic population untested. Second, DOT image reconstruction (recon) is complex and time-consuming, incompatible with the need for timely clinical decision- making. This project aims to address these translational barriers by developing and validating a two-pronged

DeepTOBIDx approach that leverages, on the one hand, a seamless integration between high-density DOT and diagnostic spot compression for lesion-targeted DBT-DOT imaging, and the other, a novel multimodal DNN model to achieve unprecedented image quality with no human-in-the-loop. From the hardware aspect (Aim 1),

we will engineer a pair of removable optical probes that house more source and detector optodes to achieve seamless integration of high-density DOT with the DBT spot compression paddle for high-resolution imaging on targeted lesions. From the image recon aspect (Aim 2), we will develop a novel multimodal DNN to directly map

sensor-domain DOT data to the image domain and further leverage the anatomical DBT to instantaneously obtain optical images of unprecedented quality. The synthetic-to-real domain adaptation of the DNN model is adequately addressed by using VICTRE and patient-derived anthropomorphic digital phantoms to represent

complex breast anatomy and lesion characteristics and by adding a realistic noise profile of the imaging system modeled by a generative adversarial network from real measurements. Finally, in Aim 3, the clinical value of the DeepTOBIDx approach will be assessed in a rigorously designed blinded multi-reader study on a 210-patient

prospective cohort to determine if the joint interpretation of clinical and DOT images can effectively reduce the number of unnecessary biopsies. If validated successfully, DeepTOBIDx can facilitate the integration of DOT with diagnostic mammography in clinical practice and result in direct benefit to breast cancer patients by avoiding

unnecessary procedures and associated burdens. The proposed deep learning-based image recon framework could also benefit broader clinical applications using other tomographic imaging modalities.