Morphological and textural descriptors analysis of digital mammograms with radiological findings to support breast...

Objective. To classify digital mammograms based on radiological findings using morphology and texture descriptors with artificial neural networks (ANN) for breast cancer detection.Approach.The mammography dataset from High Specialty Regional Hospital of Oaxaca (HRAEO) (median patient age (mpa), 48 years [interquartile range (IQR), 41-54 years]) with radiological findings was retrospectively analyzed. All patients underwent breast biopsy and were not previously treated. External testing was performed using mammograms from the National Cancer Institute (INCAN) (mpa: 47 years [IQR, 37-62 years]). The morphology was analyzed using a circularity descriptor (к), and the texture was analyzed using the mean height/width ratio of the extrema descriptor (ρ). These results were compared with cancer/benign histopathology, which was binarily classified using ANNs. The F1-score, Cohen's kappa (K), and area under the ROC curve (AUC) were employed as evaluation metrics, and the Wilcoxon rank-sum test was used for statistical analysis (h = 0, with p > 0.05, was considered as not statistically significant).Main results.216 raw mammograms from HRAEO and 33 mammograms from INCAN (95 + 16 breast cancer and 121 + 17 benign findings) were included. The best internal testing results were obtained with a one-hidden-layer ANN with 100 neurons, achieving a F1-score of 0.95, K of 0.91, and an AUC of 0.953 (95% confidence interval [CI]: 0.917, 0.977) (h = 0, p > 0.99). However, the external testing results were significantly lower: 0.38 F1-score, 0.02 K, and 0.509 AUC (95% CI: 0.344, 0.664) (h = 0, p = 0.14) due to not exactly meeting the inclusion criteria and possible demographic and spectrum bias, or domain-adaptation issues.Significance. The proposed morphology (к) and texture (ρ) descriptors show promise for detecting breast cancer in raw mammograms, with radiological findings, in a local context. However, their poor external performance highlights the need for substantial further work before this approach can be deemed suitable for broader diagnostic applications.