With 604,000 new cases and 342,000 deaths in 2020, cervical cancer is one of the most common malignancies in women which seriously threatens their physical and mental health. Fortunately, it is a cancer that can be eliminated via primary prevention strategies comprising of a human papillomavirus (HPV) vaccine, early detection, and timely treatment.
Still, the American Cancer Society recommends that vaccinated women also be screened the same as unvaccinated women because it is difficult to avert risk entirely. Screening and treatment of precancerous lesions in women are also a cost-effective way to prevent cervical cancer. Thus, routine screening for cervical cancer is still important to women. Two cervical cancer screening methods that are widely used are the Papanicolaou test (Pap test) and the HPV DNA test. However, accurate diagnosis remains difficult owing to various factors.
To tackle this challenge, a team of researchers led by Rebecca Richards-Kortum, PhD, the Rice University Malcolm Gillis University Professor of Bioengineering and Electrical and Computer Engineering, and Director of the Rice 360[degrees] Institute for Global Health, developed a deep learning-based computer-aided diagnostic (CAD) system to interpret high-resolution images and detect precancer and cancer as an alternative to conventional biopsy and histopathology. The study was published in Computerized Medical Imaging and Graphics (2022; https://doi.org/10.1016/j.compmedimag.2022.102052).
In recent years, artificial intelligence (AI)-based medical diagnostic applications have been on the rise in various diseases. Here, Richards-Kortum and colleagues investigated whether low-cost imaging technologies combined with a deep learning-based CAD system could serve as low-cost alternatives to diagnose cervical precursors in vivo and optimize cervical cancer prevention programs. These precursors, also known as cervical intraepithelial neoplasia (CIN), are graded as CIN 1-3 based on increasing severity.
In brief, the team describes a compact multi-task convolutional neural network (CNN) framework that uses nuclear segmentation to diagnose CIN 2 or more severe from a high-resolution endomicroscope (HRME). The network also incorporates patient HPV status as an additional clinical attribute to inform prediction.
A dataset of HRME images from two large diagnostic studies was used to train (n=959), validate (n=333), and test (n=335) the algorithm. An external evaluation set with images from 508 patients was used to further validate the findings. The results demonstrated that the proposed multi-task network for cervical precancer and cancer diagnosis consistently outperformed classification based on morphologic features, as well as state-of-the-art deep learning benchmarks, even as the volume of training data was reduced.
The strategy achieved a test per patient area under the receiver operating characteristic curve (AUC-ROC) of 0.87. The diagnostic performance was on par with expert colposcopic impression, with a test sensitivity and specificity of 0.94 (P=.3) and 0.58 (P=1.0), respectively. Additionally, as HPV infections are at a higher risk of developing cervical cancer, incorporating HPV status as a diagnostic feature to inform prediction boosted the per patient improved test specificity to 0.71 at a sensitivity of 0.94.
Oncology Times connected with Richards-Kortum for her perspectives on computer-aided diagnosis in cervical cancer. Her research focuses on improving early detection of cancers and other diseases, especially in low-resource settings.
Oncology Times: What are some of the factors for why accurate diagnosis of cervical cancer remains difficult?
Richards-Kortum: "Human papillomavirus (HPV) vaccination, along with detection and treatment of cervical precancerous lesions, has been highly effective at preventing cervical cancer in high-income countries. In these settings, patients who screen positive by cytology or HPV testing are referred to a second visit for colposcopy and biopsy of clinically suspicious lesions. Patients with pathologically high-grade lesions then require a third visit for treatment. This three-visit process for screening, diagnosis, and treatment of cervical precancers is difficult to implement in low- and middle-income settings due to a scarcity of trained professionals, lack of affordable equipment, lack of pathology services, and high rates of loss to follow-up."
Oncology Times: Machine learning, deep convolutional neural networks, and AI algorithms can revolutionize human health. How can the application of CAD systems in the early screening and diagnosis of cervical cancer be conducive to addressing limited human resources and improving diagnostic accuracy?
Richards-Kortum: "In low- and middle-income settings, the availability of colposcopy and pathology services is often limited. Low-cost imaging technologies coupled with deep learning-based CAD systems could address the challenges faced by low- and middle-income countries in expanding and optimizing cervical cancer prevention programs. Several CAD systems have already been developed to automate interpretation of colposcopic images, and algorithms that leverage advances in deep learning have shown improved diagnostic performance. Artificial intelligence systems could help streamline care, enabling rapid and accurate treatment decisions at a low cost."
Oncology Times: Where do you see your deep learning-based CAD system fitting within the diagnostic workup of cervical cancer?
Richards-Kortum: "We developed our deep learning-based CAD system to interpret high-resolution microendoscopy images and detect precancer and cancer. In this study, we showed that our system performed comparable to expert colposcopy for cervical precancer and cancer diagnosis, demonstrating its potential to aid clinical decision-making in real time, reducing the need for clinical resources and multiple patient visits."
Oncology Times: What limitations of the current study remain to be addressed for this system to be integrated into clinical settings?
Richards-Kortum: "Currently our system relies on clinical guidance for placement of the imaging probe. When an expert colposcopist is unavailable, this presents a limitation. Advancement in the field of colposcopy image analysis may aid in guiding appropriate placement of the imaging probe by localizing high-risk areas on the cervix."
Dibash Kumar Das is a contributing writer.