Cloud-Based AutoML Classification of Malignant and Non-malignant Colorectal Findings on Abdominal Computed Tomography: A Proof-of-Concept Study.
Authors
Affiliations (7)
Affiliations (7)
- Internal Medicine, California Northstate University College of Medicine, Elk Grove, USA.
- Osteopathic Medicine, Touro University California, Vallejo, USA.
- Internal Medicine, California Health Sciences University, Fresno, USA.
- Medicine, California University of Science and Medicine, Colton, USA.
- Medicine, The University of Oklahoma, Norman, USA.
- Medicine, University of California, Los Angeles, Los Angeles, USA.
- Internal Medicine, East Tennessee State University Quillen College of Medicine, Johnson City, USA.
Abstract
Colorectal cancer (CRC) is a major cause of cancer-related morbidity and mortality worldwide, and computed tomography (CT) plays an important role in its clinical evaluation and staging. However, CT interpretation may be influenced by lesion conspicuity, acquisition characteristics, and reader experience. Artificial intelligence (AI)-based image analysis may provide a reproducible adjunct to conventional radiologic assessment. This study evaluated a cloud-based automated machine-learning image-classification model for distinguishing malignant from non-malignant abdominal CT images. Images were obtained from the publicly available ColonCancerCT-2025 dataset, which contains CT imaging and clinical information from 40 patients, including 12 malignant and 28 non-malignant cases. For the present analysis, 3,827 CT images were included. The images were divided into training, validation, and test sets in an approximately 80:10:10 ratio, comprising 3,062 training images, 383 validation images, and 382 test images. Patient-level separation among these partitions could not be confirmed from the retained study workflow. A single-label binary image-classification model was developed using Google Cloud AutoML (Google LLC, Mountain View, CA, USA). Model performance was evaluated using average precision, precision, recall, precision-recall analysis, confidence-threshold behavior, and a normalized confusion matrix. A confidence threshold of 0.50 was used for the primary performance assessment. At this threshold, the classifier achieved an overall average precision of 0.998, with precision and recall of approximately 99%. Class-specific average precision was 0.998 for malignant images and 0.999 for non-malignant images. The normalized confusion matrix demonstrated correct classification of 97% of malignant images and 100% of non-malignant images. Representative correctly classified images received prediction probabilities of 0.947 for the malignant class and 0.960 for the non-malignant class. These findings demonstrate the feasibility of cloud-based automated image classification for colorectal CT analysis; however, the limited number of patients, lack of confirmed patient-level separation between dataset partitions, absence of independent external testing, and image-level rather than examination-level evaluation substantially limit conclusions regarding clinical generalizability. Larger patient-level, multi-institutional studies are required before such a model can be considered for clinical decision support.