An AI-powered, lightweight App offers hope for early-stage cervical cancer diagnosis

An AI-powered, lightweight, diagnostic, web-based App — CerviSpectraDiag — developed by researchers from IIT Kharagpur and KIIT University for early detection of cervical cancer has an accuracy of 84.73%. It offers diagnosis in near real time while maintaining privacy and data security as processing of spectral data of cervical tissue happens at the device level

An AI-powered, lightweight, diagnostic, web-based App — CerviSpectraDiag — developed by researchers from IIT Kharagpur and KIIT University in Bhubaneswar, Odisha for early detection of cervical cancer has the potential of serving women in rural and hard-to-reach areas. With an accuracy of 84.73%, the CerviSpectraDiag web-based App is capable of offering diagnosis in near real time while maintaining privacy and data security. Results of the proof-of-concept study were published recently in IEEE Transactions on Computational Social Systems.

According to a July 2024 paper in the journal Reproductive Health, in 2020, there were an estimated 6,04,127 cases and 3,41,831 deaths attributed to cervical cancer globally. Of these, India accounted for nearly one-fifth of new cases and about one-fourth of deaths. Cervical cancer was the second most common cancer in India in both incidence (18.3%) and cancer mortality (18.7%) among women in 2020, with a 5-year prevalence of 18.8%.

Despite the high number of cervical cancer cases and deaths each year in India, at just 2%, the prevalence of cervical cancer screening is very low; it ranges from 0.2% in West Bengal and Assam to 10.1% in Tamil Nadu, as per a June 2023 study in the journal BMC Women’s Health. There is also an urban-rural divide in screening with higher screening coverage of 2.4% in urban areas and 1.8% in rural areas. Educational inequalities, low awareness and affordability are some of the main reasons for low uptake of screening in India. Traditional biopsies, though effective, are often painful, expensive, and time-consuming. CerviSpectraDiag addresses these challenges by offering quick diagnosis and at less cost and discomfort.

Local network nodes reduce data leakage

CerviSpectraDiag is a lightweight web-based App that has relied on personalised federated learning to train the model using multiple decentralised, low-resource devices such as smartphones and laptops with just 4GB RAM and dual core CPUs for machine learning. Diagnosis of cervical cancer is based on spectral data obtained from cervical tissue, which can be uploaded to local devices for processing. “At the heart of the framework lies CerviSpectraYOLO, a modified version of the popular YOLO deep learning model. It has been enhanced through personalised federated learning to process spectral data of cervical tissue without transferring data to central servers,” says Sabyasachi Mukhopadhyay, a PhD student at the Centre for Computational & Data Sciences, IIT Kharagpur and the first author of the paper.

The processing and interpretation of the spectral data, as well as the generation of diagnostic results, are performed locally on devices, with only the aggregated results or model updates sent to the central server. “Local processing at the device level, without transmitting raw data to the cloud or a central server, greatly minimises the risk of data leakage,” says Mukhopadhyay.

High accuracy and easy accessibility

“When tested across multiple local nodes, the system achieved an impressive 84.73% top-1 accuracy, outperforming other lightweight YOLO architectures,” he says. “The study highlights how this lightweight tool can deliver accurate, explainable, and rapid — even on low-resource devices like smartphones.”

Capable of working on low-resource devices with just 4GB RAM and dual-core CPUs, CerviSpectraDiag is particularly suited for resource-limited settings. “Relying only on low-resource devices makes the web-based App an ideal candidate for deployment in rural and semi-urban areas, where access to specialised healthcare facilities is scarce,” Mukhopadhyay explains. “The App offers an alternative to standard biopsies, which is both expensive, invasive and time-consuming, and hence not accessible by many women.”

The App diagnoses and classifies cervical cancer either as healthy or as precancerous. In precancerous lesions of cervical cancer, the App can classify them into grades — grade I, grade II, and grade III — based on the severity of the cellular changes. For deep learning and training, spectral data were collected from GSVM Medical College, Kanpur, comprising 15 samples of normal tissue, 15 samples of Grade I, nine samples of Grade II, and 18 samples of Grade III precancerous tissues.

Based on these original dataset, standard data augmentation methods were followed to increase the spectral samples of both healthy and precancerous lesions such that the centralised spectral database contains a total of 58,311 spectral samples. Of the total number of spectral samples, 15,345 spectral samples are of healthy or normal tissues, 15,345 spectral samples are of grade I, 9,207 spectral samples are of grade II, and 18,414 spectral samples are of grade III cervical precancerous tissues.

“To make the data processing more efficient, we followed standard data preprocessing techniques to clean and normalise the data,” says Mukhopadhyay. “80% of the total dataset (58,311 spectral samples) was used for training the algorithm, while the remaining 20% data were used for testing purposes. A portion of the training data was further set aside for validation,” says Mukhopadhyay.

Patient-friendly explanation of diagnosis

The second core component that complements the diagnostic engine is CerviSpectraLangChain, which uses advanced language models to generate clear, patient-friendly explanations of diagnosis in both English and Hindi. “This ensures that complex medical insights are communicated effectively to both patients and doctors alike,” he says.

The researchers also developed a user-friendly web app with separate dashboards for doctors and patients. Physicians can upload patient data, review AI-generated reports, and provide treatment guidance, while patients can securely access their results, book appointments, and communicate with healthcare providers — all through a secure interface, he explains.

Next, the researchers plan to integrate the system with electronic health records for continuous treatment support, moving closer to a cost-effective, scalable healthcare solution that could significantly reduce the global burden of cervical cancer.

Featured image credit: National Cancer Institute