Browsing by Subject "YOLOv8"

Now showing 1 - 1 of 1
  • Thumbnail Image
    Item
    Detección automática de masas en mamografías digitales mediante técnicas de inteligencia artificial
    (Pontificia Universidad Javeriana Cali, 2025) García Gallego, Fabián Antoyne; Palta, Felipe
    Early detection of breast cancer through mammography is a problem of high clinical impact; however, manual localization of masses can be laborious due to low contrast, tissue overlap, and high morphological variability of lesions. This work presents an automatic mass detection system in digitized mammograms using the one-stage YOLO family of detectors. The mass subset of the CBIS-DDSM dataset was used to build a preprocessing pipeline that (i) loads mammograms in DICOM format, (ii) uses ROI masks to derive consistent bounding boxes, and (iii) generates annotations in YOLO format and training, validation, and test partitions. A pre-trained YOLOv8m model was selected as the base model and fine-tuned via transfer learning to a single class (mass) with imgsz=640. On the validation set, the model achieved P = 0,586, R = 0,544, mAP@0,5 = 0,533, and mAP@0,5:0,95 = 0,243. For a more clinically interpretable evaluation, an inference pipeline with post-processing based on Extra NMS was implemented to remove highly overlapping duplicate detections, and two complementary metrics were computed: (i) the Jaccard Index (IoU) to quantify spatial agreement between predictions and ground truth, and (ii) an FROC curve (sensitivity vs. FPPI) to analyze the trade-off between sensitivity and false positives per image. On the test set, 39,61 % of the images exhibited IoU ≥ 0,60 (143 of 361), and the FROC analysis reported sensitivities of 0.5435 at FPPI ≤ 0,5, 0.6332 at FPPI ≤ 1,0, and 0.6860 at FPPI ≤ 2,0. These results suggest that the proposed approach can localize masses with significant spatial concordance in a relevant fraction of cases, providing a reproducible basis for clinical decision support systems.