Editors: Liron Pantanowitz, MD, PhD, MHA, chair of the Department of Pathology and professor of pathology, University of Pittsburgh Medical Center, and Matthew G. Hanna, MD, vice chair of pathology informatics and associate professor, Department of Pathology, University of Pittsburgh Medical Center.
Validation of AI-assisted Pap test screening using a digital diagnostic system
October 2024—The authors conducted a study to evaluate the Genius digital diagnostic system compared with manual light microscopy diagnosis using ThinPrep Pap test slides. Six cytologists and three cytopathologists participated in the study. They received 1.5 days of training on the Genius digital system from the manufacturer (Hologic, Marlborough, Mass.). They then analyzed 319 ThinPrep Pap test cases in the authors’ institutional cytology archive that represented a range of Bethesda System categories typically encountered in routine practice. The study participants assessed diagnostic accuracy by comparing digital and manual results to the original Pap test diagnosis, which was considered the reference diagnosis, or ground truth. They evaluated the accuracy across eight diagnostic categories—unsatisfactory; negative for intraepithelial lesion or malignancy; atypical squamous cells of undetermined significance; low-grade squamous intraepithelial lesions; high-grade squamous intraepithelial lesions; atypical squamous cells, cannot exclude HSIL; atypical glandular cells; and malignant. The participants also assessed accuracy using condensed three- and four-category classification groups based on the Bethesda System and clinical management. The results showed that digital slide review had significantly higher concordance with the ground truth compared with manual microscopy for all classifications. The cytologist and cytopathologist study participants also had higher accuracy rates with digital slide review and reduced review time (3.2 minutes) than with manual microscopy (5.9 minutes). Human papillomavirus status was considered for atypical cases, but diagnostic concordance between HPV+ and HPV− cases did not differ significantly between the methods. The authors concluded that the digital system improves diagnostic accuracy and provides a statistically significant reduction in review time, making it more reliable and efficient than manual microscopy.
Cantley RL, Jing X, Smola B, et al. Validation of AI-assisted ThinPrep Pap test screening using the Genius digital diagnostic system. J Pathol Inform. 2024;15. doi.org/10.1016/j.jpi.2024.100391
Correspondence: Dr. Richard L. Cantley at rcantley@med.umich.edu
Use of artificial intelligence for gunshot wound classification
In forensic medicine, gunshot wounds are classified as entrance or exit wounds, based on specific characteristics. The wounds are often photographed for documentation purposes. Forensic pathologists and emergency room physicians consider such factors as muzzle imprints, peripheral tears, stippling, bone beveling, and wound border irregularity when differentiating between entrance and exit wounds. However, not every case is straightforward, and some wounds can be challenging to decipher. The authors of this study developed a deep learning model to classify entry and exit wounds on digital color images. They collected 2,418 digital color images of entrance (n = 2,028) and exit (n = 1,314) wounds to train a ConvNext Tiny deep learning model. The artificial intelligence (AI)-based model achieved accuracy of 87.99 percent, precision of 83.99 percent, specificity of 88.19 percent, F1-score of 85.81 percent, and area under the curve of 0.946 on a holdout test set. It was able to correctly classify 88.19 percent of entrance wounds and 87.71 percent of exit wounds. The results were comparable to what a forensic pathologist could achieve without other morphologic cues. Two experienced forensic pathologists reviewed the 100 images that the model had the most difficulty classifying, and both pathologists outperformed AI. The authors concluded that their study demonstrates that a deep learning model can discern gunshot wounds in digital images, aiding forensic pathology.
Cheng J, Schmidt C, Wilson A, et al. Artificial intelligence for human gunshot wound classification. J Pathol Inform. 2023;15. doi.org/10.1016/j.jpi.2023.100361
Correspondence: Dr. Jerome Cheng at jeromech@med.umich.edu