With an increase in deep learning-based methods, the call for explainability of such methods grows, especially in high-stakes decision making areas such as medical image analysis. This survey presents an overview of explainable artificial intelligence (XAI) used in deep learning-based medical image analysis. A framework of XAI criteria is introduced to classify deep learning-based medical image analysis methods. Papers on XAI techniques in medical image analysis are then surveyed and categorized according to the framework and according to anatomical location. The paper concludes with an outlook of future opportunities for XAI in medical image analysis.

In women with extremely dense breasts, a deep learning model identified breast MRI examinations without cancer with high certainty, dismissing nearly 40% of normal MRI examinations without missing ... Background Supplemental screening with MRI has proved beneficial in women with extremely dense breasts. Most MRI examinations show normal anatomic and physiologic variation that may not require rad...

Objectives: To reduce the number of false-positive diagnoses in the screening of women with extremely dense breasts using magnetic resonance imaging (MRI), we aimed to predict which BI-RADS 3 and BI-RADS 4 lesions are benign. For this purpose, we use computer-aided diagnosis (CAD) based on multiparametric assessment. Materials and Methods: Consecutive data were used from the first screening round of the DENSE (Dense Tissue and Early Breast Neoplasm Screening) trial. In this trial, asymptomatic women with a negative screening mammography and extremely dense breasts were screened using multiparametric MRI. In total, 4783 women, aged 50 to 75 years, enrolled and were screened in 8 participating hospitals between December 2011 and January 2016. In total, 525 lesions in 454 women were given a BI-RADS 3 (n = 202), 4 (n = 304), or 5 score (n = 19). Of these lesions, 444 were benign and 81 were malignant on histologic examination. The MRI protocol consisted of 5 different MRI sequences: T1-weighted imaging without fat suppression, diffusion-weighted imaging, T1-weighted contrast-enhanced images at high spatial resolution, T1-weighted contrast-enhanced images at high temporal resolution, and T2-weighted imaging. A machine-learning method was developed to predict, without deterioration of sensitivity, which of the BI-RADS 3– and BI-RADS 4–scored lesions are actually benign and could be prevented from being recalled. BI-RADS 5 lesions were only used for training, because the gain in preventing false-positive diagnoses is expected to be low in this group. The CAD consists of 2 stages: feature extraction and lesion classification. Two groups of features were extracted: the first based on all multiparametric sequences, the second based only on sequences that are typically used in abbreviated MRI protocols. In the first group, 49 features were used as candidate predictors: 46 were automatically calculated from the MRI scans, supplemented with 3 clinical features (age, body mass index, and BI-RADS score). In the second group, 36 image features and the same 3 clinical features were used. Each group was considered separately in a machine-learning model to differentiate between benign and malignant lesions. We developed a Ridge regression model using 10-fold cross validation. Performance of the models was analyzed using an accuracy measure curve and receiver-operating characteristic analysis. Results: Of the total number of BI-RADS 3 and BI-RADS 4 lesions referred to additional MRI or biopsy, 425/487 (87.3%) were false-positive. The full multiparametric model classified 176 (41.5%) and the abbreviated-protocol model classified 111 (26.2%) of the 425 false-positive BI-RADS 3– and BI-RADS 4–scored lesions as benign without missing a malignant lesion. If the full multiparametric CAD had been used to aid in referral, recall for biopsy or repeat MRI could have been reduced from 425/487 (87.3%) to 311/487 (63.9%) lesions. For the abbreviated protocol, it could have been 376/487 (77.2%). Conclusions: Dedicated multiparametric CAD of breast MRI for BI-RADS 3 and 4 lesions in screening of women with extremely dense breasts has the potential to reduce false-positive diagnoses and consequently to reduce the number of biopsies without missing cancers.

Objectives To describe and to correlate tumor characteristics on multiparametric 7 tesla (T) breast magnetic resonance imaging (MRI) with prognostic characteristics from postoperative histopathology in patients with breast cancer. Materials and Methods Institutional review board approval and written informed consent of 15 women (46–70 years) with 17 malignant lesions were obtained. In this prospective study (March 2013 to March 2014), women were preoperatively scanned using dynamic contrast-enhanced MRI, diffusion-weighted imaging, and 31-phosphorus spectroscopy (³¹P-MRS). The value of the protocol was assessed to quantify tumor differentiation and proliferation. Dynamic contrast-enhanced MRI was assessed according to the American College of Radiology Breast Imaging Reporting and Data System-MRI lexicon. Apparent diffusion coefficients (ADCs) were calculated from diffusion-weighted imaging. On ³¹P-MRS, at the location of the tumor, the amount of phosphorus components was obtained in a localized spectrum. In this spectrum, the height of phosphodiester (PDE) and phosphomonoester (PME) peaks was assessed to serve as a measure for metabolic activity, stratifying tumors into a PDE > PME, PDE = PME, or PDE < PME group. Tumor grade and mitotic count from resection specimen were compared with the MRI characteristics using explorative analyses. Results On dynamic contrast-enhanced MRI, the mean tumor size was 24 mm (range, 6–55 mm). An inverse trend was seen between ADC and tumor grade (P = 0.083), with mean ADC of 867 × 10⁻⁶ mm²/s for grade 1 (N = 4), 751 × 10⁻⁶ mm²/s for grade 2 (N = 6), and 659 × 10⁻⁶ mm²/s for grade 3 (N = 2) tumors. Between ³¹P-MR spectra and mitotic count, a relative increase of PME over PDE showed significant association with increasing mitotic counts (P = 0.02); a mean mitotic count of 6 was found in the PDE greater than PME group (N = 7), 8 in the PDE = PME group (N = 1), and 17 in the PDE < PME group (N = 3). Conclusions Multiparametric 7 T breast MRI is feasible in clinical setting and shows association between ADC and tumor grade, and between ³¹P-MRS and mitotic count.