Oral cancer is a common cancer that affects the head and neck region. Around 50% of oral cancer patients do not survive for more than five years even after treatment as many patients seek medical help only during the late stages of the disease. By the time the patients seek help, the cancer may require advanced treatment, and may even have spread to other areas.
Oral cancer often develops from white patches clinically known as oral leukoplakia (OL) and oral lichenoid mucositis (OLM). These white patches may appear long before a diagnosis of oral cancer, and their early detection and continuous monitoring are crucial to prevent the development of cancer. However, it is sometimes challenging to predict which OL/OLM lesions will develop into oral cancers as the global risk rates of progression from OL/OLM to cancer vary from 0.4 to 40.8%.
Routine visitation and multiple biopsies are often scheduled during OL/OLM surveillance. In addition, patient monitoring may continue for many years leading to fatigue and lack of compliance with hospital appointments. Therefore, it will be better to determine the risk of cancer development in these patients on an individual basis to allow health professionals to use this information to formulate specific treatment and follow-up schedules for each patient.
Researchers from the Faculty of Dentistry and the Li Ka Shing Faculty of Medicine, the University of Hong Kong (HKU); Department of Pathology, Queen Mary Hospital, HK; and College of Medicine and Dentistry, James Cook University, Queensland collaborated to develop a web platform that can be applied to automatically generate an individualized prediction of the risk of oral cavity cancer occurrence in those with OL or OLM for up to 20 years following diagnosis.
The results are published in the journal Cancers in an article entitled “Deep learning predicts the malignant-transformation-free survival of oral potentially malignant disorders”.
The freely available web tool, based on the artificial intelligence algorithm ‘DeepSurv’ was trained and tested with data from patients with OL/OLM treated in Hong Kong (716 patients) and Newcastle Upon Tyne, UK (382 patients). As the patients have already been under review for many years, their true risk levels were already known, and the study showed that the artificial intelligent model was able to accurately predict their risk levels at different time points during their follow-up hospital visits.
The DeepSurv algorithm was selected due to its superior performance for the use of routine demographic, clinical, pathological, and treatment information of these patients for cancer risk prediction following a series of validation exercises.
On a validation subset of the Hong Kong cohort, ‘DeepSurv’ was able to predict the correct cancer risk level for 95% of the cases. This was according to an integrated Brier score of 0.04, with a score below 0.25 generally depicting a tool that may be useful in real-world applications.
The algorithm is further able to generate correct risk levels for 82% of the patients in the British cohort which suggests its utility in other populations as well.
To expatiate on how the interactive web tool functions, it requires 26 pieces of information on the demography, clinical and pathological description of the disease, and treatment received by the OL/OLM patient. The predicted output from the web tool includes a curve from which the different risk levels (vertical axis) can be visualized at each time point (horizontal axis). These predicted risk levels have been shown to be accurate up to 17 years from the time that the information was entered.
This will assist health professionals in the selection and prioritization of treatment strategies and close-monitoring schedules for high-risk patients, especially in resource-limited hospitals. Ultimately, this is expected to improve on the currently available methods of prevention and early diagnosis of oral cancer.
The prediction curve may also be used for individual cancer risk estimation and inform health professionals when to commence very close monitoring of patients when a certain risk level is reached. (Figure 1)
For OL/OLM patients, risk awareness may motivate them to regularly attend routine follow-up visits and allow them to make informed decisions when providing consent for biopsy when required.
Of note, the predicted risk curve may change with varying input data such as smoking and alcohol drinking status, parts of the mouth that are affected, treatment received, lesion recurrence, and the severity of epithelial dysplasia during treatment monitoring.
“While the web tool has been found promising based on our validation exercises, users should know that it is still primarily a research-based tool and requires further prospective optimization,” said Dr Richard Su, Clinical Associate Professor in Oral and Maxillofacial Surgery (OMFS), Faculty of Dentistry, HKU, who led the study.
“Since cancer development involves many alterations at the molecular level that may occur before disease diagnosis, we will in the future optimize the web tool by including information on molecular biomarkers for cancer development in OL and OLM.” Dr Su added. It is expected that the inclusion of the information in the web tool will improve the precision of the predicted risk estimates. The updated web tool will then be evaluated for its clinical efficacy and its impact in the care of OL and OLM in a clinical trial.
The research team
The research team was led by Dr Richard Su, Clinical Associate Professor in Oral and Maxillofacial Surgery (OMFS), with co-investigators including PhD student Dr John Adeoye, Research Assistant Professor of OMFS Dr Siu-Wai Choi, Clinical Associate Professor in Oral Medicine Dr Li-Wu Zheng and Research Assistant Professor in Clinical Artificial Intelligence Dr Mohamad Koohi-Moghadam, HKU Faculty of Dentistry; and Dr Anthony Lo of the Department of Pathology, Queen Mary Hospital, HK; Dr Raymond Tsang and Dr Velda Chow, Department of Surgery, Li Ka Shing Faculty of Medicine, HKU and Professor Peter Thomson, College of Medicine and Dentistry, James Cook University, Queensland.
‘DeepSurv’ is available for free:
The Journal “Deep learning predicts the malignant-transformation-free survival of oral potentially malignant disorders”:
For media enquiries:Ms Melody Tang, Senior Communications Officer of the Faculty of Dentistry, HKU, Tel: 28590494, Email:
melodytang@hku.hk.
口腔癌是一種頗常見的癌症，影響頭頸部位。大約五成的口腔癌患者存活不超過五年，主要是發現時已屬較後期，需要複雜的治療方案，癌細胞亦有可能已擴散到其他部位。
口腔癌初起的表徵很多時是口腔内一些白斑，臨床診斷為口腔白斑病（oral leukoplakia OL） 和口腔類扁平苔蘚黏膜炎（oral lichenoid mucositis OLM)。這些白斑可能在出現一段長時間後才被確診口腔癌，所以能儘早發現並持續監察，防止它們惡化發展至為重要。
然而，OL/OLM轉化為癌症，目前全球的風險比率由0.4% 至40.8%不等， 因此要得知那些白斑最終會演變成口腔癌，臨床上是極大的挑戰。同時，病人一般需要定期覆診和接受多次切片檢查，持續監察多年，令患者感到疲憊甚至拒絕覆診。因此，若能因應個別病人的狀況，預測其演化成癌症的風險，將有助醫療人員為病人制定特定的醫療策略和跟進方式。
來自香港大學（港大）牙醫學院、李嘉誠醫學院，瑪麗醫院病理部，以及昆士蘭詹姆士庫克大學醫學與牙科學院的研究人員，共同研發了一個網上平台，在輸入相關資料後，能針對患有OL或OLM的病人，預測他們的口腔癌病變風險，預測期長達20年之久。網上平台開放給醫護人員和公衆免費使用。
研究結果已於學術期刊《Cancers》發表，文章題爲 「深度學習能預測口腔癌前病變患者的無惡性轉化存活期」（“Deep learning predicts the malignant-transformation-free survival of oral potentially malignant disorders”）。
研究團隊採用「DeepSurv」人工智能演算法，此演算法預測癌症病變的表現優異，經一連串的驗證練習後，能根據病人的一般個人統計資料、臨床和病理數據，以及治療資料等準確預測病人罹患的風險。他們利用香港和英國紐卡素泰恩的OL和OLM的病人數據，訓練人工智能平台並測試其準確度。由於這些病人已被跟進多年，其病歷發展已經明朗，可與平台的預測對照。測試證實，新平台能準確推算這些病人的實際病變發展，在他們每個復診的時間點，平台預測的癌症風險水平與實際情況相吻合。
在香港用作驗證的716名病人的數據，「DeepSurv」能正確預測95% 病人的癌症風險水平。在反映預測準確度的Brier綜合得分是0.04。通常預測工具如果得分低於0.25，已可轉化作實際臨床應用。
而利用英國紐卡素泰恩的382名病人數據，平台正確預測82%病人的風險水平，表明其對不同的人口也具實用性。
團隊期望這人工智能工具有助改善口腔癌的預防和早期診斷。利用開放平台，醫護人員可為高危患者制訂監測時間表和治療策略，對資源有限的醫院，有助其確定處理病患的優先次序。
使用平台預測病變風險，需輸入二十六項有關病人的個人背景資料、病症的臨床和病理描述，以及接受的治療等。就每個個案的風險評估，平台會呈現一條曲線，展示每個時間點預測的風險級別。而當患者達到一定的風險水平時，醫護人員可開始加密對患者的監察。（圖一）
除了醫護人員，OL/ OLM患者使用平台可了解自己的病情趨勢，這有助提高風險意識，鼓勵他們定期覆診，並在有需要決定是否做切片組織等進一步檢查時，得以參考。
而預測的風險水平，從輸入資訊日起計17年内維持準確度。值得注意的是，預測的風險曲線會隨著輸入數據的轉變而有所變化，例如吸煙和飲酒狀況、受影響的口腔部位、接受的治療、復發情況以及治療監測期間口腔上皮變異的程度等，因此輸入數據需適時更新。
帶領研究的港大牙醫學院口腔頜面外科臨床副教授蘇宇雄醫生說: 「雖然驗證練習的結果，證明這人工智能預測工具非常可靠，但用者要知道，它的開發主要以研究為基礎，故仍有需要再進一步優化發展。」
「由於癌症的發展，在確診前涉及很多生物分子層面上的轉變，我們計劃加入OL和OLM演變成癌症的過程中相關分子生物標記的訊息，優化預測平台，提高風險估計的精準度，之後再評估其臨床成效，以及在臨床試驗中了解其對OL和OLM護理的影響。」蘇醫生補充說。
研究團隊
研究團隊由香港大學牙醫學院口腔頜面外科臨床副教授蘇宇雄醫生帶領，成員包括博士生John Adeoye醫生、研究助理教授蔡小慧博士、口腔黏膜科學臨床副教授鄭立武醫生、臨床人工智能研究助理教授Koohi-Moghadam Mohamad博士，香港瑪麗醫院病理部羅頴業醫生，香港大學醫學院外科學系曾敬賢醫生和周令宇醫生，以及昆士蘭詹姆士庫克大學醫學與牙科學院湯迅教授。
‘DeepSurv’ 免費提供:
期刊「深度學習能預測口腔癌前病變患者的無惡性轉化存活期」(Deep learning predicts the malignant-transformation-free survival of oral potentially malignant disorders)
傳媒查詢 :
香港大學牙醫學院 高級傳訊主任 鄧慧中 (Melody Tang)
電話︰2859 0494 / 9155 0980 / 電郵：
melodytang@hku.hk
