This work aims to implement an automated data-driven model for breast cancer detection in mammograms to support physi-cians’ decision process within a breast cancer screening or detection program. The public available CBIS-DDSM and the INbreast datasets were used as sources to implement the transfer learning technique on full-field digital mammography pro-prietary dataset. The proprietary dataset reflects a real heterogeneous case study, consisting of 190 masses, 46 asymmetries, and 71 distortions. Several Yolo architectures were compared, including YoloV3, YoloV5, and YoloV5-Transformer. In addition, Eigen-CAM was implemented for model introspection and outputs explanation by highlighting all the suspicious regions of interest within the mammogram. The small YoloV5 model resulted in the best developed solution obtaining an mAP of 0.621 on proprietary dataset. The saliency maps computed via Eigen-CAM have proven capable solution reporting all regions of interest also on incorrect prediction scenarios. In particular, Eigen-CAM produces a substantial reduction in the incidence of false negatives, although accompanied by an increase in false positives. Despite the presence of hard-to-recognize anomalies such as asymmetries and distortions on the proprietary dataset, the trained model showed encouraging detection capabilities. The combination of Yolo predictions and the generated saliency maps represent two complementary outputs for the reduction of false negatives. Nevertheless, it is imperative to regard these outputs as qualitative tools that invariably necessitate clinical radiologic evaluation. In this view, the model represents a trusted predictive system to support cognitive and decision-making, encouraging its integration into real clinical practice.

 

Introduction

Breast cancer is the most common worldwide tumor in the female population. Previous randomized trials and incidence-based mortality studies have demonstrated a sig-nificant reduction in breast cancer mortality associated with participation in breast screening programs. However, the problem of false positives and false negatives persists as a concern. Most of these errors can be attributed to dense breasts (masking effect), as well as human factors such as radiologist perception and erroneous decision-making behaviors. Additionally, the inherent imaging characteristics of tumors contribute to the issue, with benign masses often resembling malignant ones and malignant masses sometimes mimicking benign ones. During the breast cancer diag-nosis process, the physician aims to detect all the regions of interest (ROIs) in the whole mammogram: masses, calcifi-cations, distortions, etc. Detection in the early stage of the disease is critical for planning new examinations, therapies, or lines of intervention. A missed detection, on the other hand, may result in irreversible injury to the patient. For this reason, breast cancer detection is the most complicated but also the most important task. Unfortunately, several pro-posed solutions in the literature do not aim to analyze the entire image, but rather limit detection to patch classifica-tion: the ROIs are first manually selected and cropped, and then the classifiers are trained to distinguish the crops. How-ever, to support and imitate the physician’s diagnostic pro-cess, an architecture capable of detecting all ROIs within the whole mammogram is required. Faster R-CNN, RetinaNet, and Yolo have encouraged the development of systems for breast cancer detection. These frameworks certainly introduce two main difficulties:

(1) the models have to learn the features of the whole mammogram, and the image resiz-ing required for training may result in the loss of critical details;

(2) since the model has to detect all ROIs among all patches of healthy tissue (i.e., non-ROIs), an unavoidable increase in the error rate must be faced. However, Yolo has proven to be an excellent tool in numerous scenarios, achiev-ing higher accuracy and inference speed rates than its object detector competitors.