DOI: https://doi.org/10.1038/s41698-024-00624-8
PMID: https://pubmed.ncbi.nlm.nih.gov/38942998
تاريخ النشر: 2024-06-28
المؤلف: Adam Shephard وآخرون
الموضوع الرئيسي: أمراض الفم وعلاجها
نظرة عامة
تشوه الظهارة الفموية (OED) هو حالة ما قبل السرطان في تجويف الفم تتميز بتباين كبير بين المراقبين وفي داخل المراقب نفسه في تصنيفها، مما لا يتنبأ بشكل متسق بالتقدم نحو الخباثة. لتحسين دقة التشخيص وقرارات العلاج، يقدم هذه الدراسة خوارزمية ذكاء اصطناعي (AI) تولد درجة خطر التحول الخبيث الفموي (OMT) من صور الشرائح الكاملة الملونة بصبغة الهيماتوكسيليين والإيوزين (H&E). تستخدم سلسلة الذكاء الاصطناعي نموذج تقسيم لتحديد وتقسيم النوى والظهارة، تليها شبكة عصبية ضحلة تستخدم ميزات شكلية ومكانية قابلة للتفسير للتنبؤ بتقدم OED.
تم التحقق من صحة الخوارزمية من خلال التحقق الداخلي المتقاطع مع مجموعة تطوير (n = 193) والتحقق الخارجي مع مجموعتين مستقلتين (n = 89). أظهرت درجة OMTscore منطقة تحت منحنى التشغيل الخاص بالمستقبل (AUROC) قدرها 0.75 واسترجاع قدره 0.92، متجاوزة أنظمة التصنيف التقليدية (ثنائي: AUROC = 0.72، استرجاع = 0.85). أشار تحليل البقاء إلى القيمة التنبؤية لدرجة OMTscore (C-index = 0.60، p = 0.02)، والتي تعادل تصنيف منظمة الصحة العالمية (WHO) (C-index = 0.64، p = 0.003) ودرجات ثنائية (C-index = 0.65، p < 0.001). من الجدير بالذكر أن التحليلات النووية كشفت عن ارتباطات كبيرة مع الخلايا اللمفاوية المحيطية والداخلية في بقع تنبؤية للحالات المتحولة (p < 0.001). تمثل هذه الدراسة أول خوارزمية مؤتمتة بالكامل وقابلة للتفسير للتنبؤ بتحول OED، مما يقدم تقدماً واعداً في الإدارة السريرية لسرطان الخلايا الحرشفية الفموية (OSCC) ويعزز جهود الكشف المبكر الضرورية لتحسين نتائج المرضى.
طرق
يستعرض قسم “الطرق” الأساليب التجريبية والتحليلية المستخدمة في الدراسة. يوضح تصميم التجارب، بما في ذلك اختيار المشاركين، والمواد المستخدمة، والإجراءات المحددة المتبعة لضمان موثوقية وصلاحية النتائج. تم إجراء التحليلات الإحصائية باستخدام برامج مناسبة لتقييم البيانات، مع تحديد مستويات الدلالة عند p < 0.05. بالإضافة إلى ذلك، يصف القسم أي نماذج رياضية أو معادلات تم تطبيقها لتفسير النتائج، مما يضمن توافق الأساليب مع أهداف البحث. يتم التأكيد على صرامة الطرق لدعم مصداقية الاستنتاجات المستخلصة من الدراسة.
نتائج
في هذه الدراسة، طورنا سلسلة متعددة الخطوات للتنبؤ بدرجة خطر OED (OMTscore) باستخدام نهج التعلم العميق. في البداية، قمنا بتدريب نموذج تعلم عميق داخلي لتقسيم الطبقات الداخلية الظهارية والنوى. تم تطبيق هذا النموذج بعد ذلك لتوليد تقسيمات عبر جميع الشرائح في مجموعاتنا.
بعد التقسيم، تم تقسيم كل شريحة إلى بلاطات أصغر، وتم استخراج ميزات شكلية على مستوى البلاطة بناءً على التقسيمات النووية داخل الظهارة. تم استخدام هذه الميزات بعد ذلك في شبكة عصبية متعددة الطبقات (MLP) للتنبؤ باحتمالية التحول إلى الخباثة، وبالتالي حساب درجة OMTscore. يبرز هذا النهج دمج تقنيات تحليل الصور المتقدمة مع النمذجة التنبؤية لتقييم خطر السرطان.
مناقشة
في هذه الدراسة، قدمنا درجة OMTscore، وهي أداة جديدة لتقييم المخاطر مصممة للتنبؤ بتحول تشوه الظهارة الفموية (OED) إلى الخباثة. بدأت سلسلة النموذج بتقسيم متزامن للحالات النووية والطبقات الداخلية الظهارية باستخدام نموذج HoVer-Net+ على صور الشرائح الكاملة الملونة بصبغة H&E. تم التحقق من أداء التقسيم على مجموعة بيانات شيفيلد الداخلية، مما أسفر عن نتائج مرضية تبرر تطبيقها على كل من المجموعات الداخلية والخارجية. بعد التقسيم، قمنا بتوليد ميزات شكلية ومكانية على مستوى البقع، والتي تم استخدامها بعد ذلك في شبكة عصبية متعددة الطبقات (MLP) لإنتاج درجة OMTscore. أظهر النموذج أداءً تنافسياً، محققاً منطقة تحت منحنى التشغيل الخاص بالمستقبل (AUROC) قدرها 0.77 في التحقق الداخلي، متجاوزاً أنظمة التصنيف التقليدية.
أكدت تحليلات البقاء بشكل إضافي على الفائدة التنبؤية لدرجة OMTscore، التي أظهرت مؤشر توافق أعلى (C-index) مقارنة بأنظمة التصنيف WHO والثنائية، خاصة في تمييز الحالات ذات المخاطر المنخفضة عن العالية. من الجدير بالذكر أن أداء درجة OMTscore كان متسقاً عبر مجموعات التحقق الخارجي، على الرغم من ملاحظة بعض التباين، على الأرجح بسبب اختلافات في خصائص العينة وقابلية تعميم نموذج التقسيم. كشفت تحليلات الميزات أن الحالات المتحولة أظهرت أعداداً أعلى من أنواع نووية معينة وتركيباً خلوياً مميزاً، مما يشير إلى أن تسلل الخلايا المناعية قد يلعب دوراً في التحول الخبيث. بشكل عام، تؤكد نتائجنا على إمكانات درجة OMTscore كأداة قوية وقابلة للتفسير للكشف المبكر عن الآفات عالية المخاطر، مع آثار لتحسين إدارة المرضى واستراتيجيات العلاج في OED. يجب أن تركز الأعمال المستقبلية على تعزيز قابلية تعميم النموذج والتحقق من صحة درجة OMTscore في مجموعات بيانات أكبر ومتعددة المراكز.
DOI: https://doi.org/10.1038/s41698-024-00624-8
PMID: https://pubmed.ncbi.nlm.nih.gov/38942998
Publication Date: 2024-06-28
Author(s): Adam Shephard et al.
Primary Topic: Oral Health Pathology and Treatment
Overview
Oral epithelial dysplasia (OED) is a premalignant condition of the oral cavity characterized by significant inter- and intra-observer variability in its grading, which does not consistently predict the progression to malignancy. To improve diagnostic accuracy and treatment decisions, this study introduces an artificial intelligence (AI) algorithm that generates an Oral Malignant Transformation (OMT) risk score from Haematoxylin and Eosin (H&E) stained whole slide images (WSIs). The AI pipeline employs a segmentation model to identify and segment nuclei and epithelium, followed by a shallow neural network that utilizes interpretable morphological and spatial features to predict OED progression.
The algorithm was validated through internal cross-validation with a development cohort (n = 193) and external validation with two independent cohorts (n = 89). The OMTscore demonstrated an area under the receiver operating characteristic curve (AUROC) of 0.75 and a recall of 0.92, surpassing traditional grading systems (Binary: AUROC = 0.72, Recall = 0.85). Survival analysis indicated the prognostic value of the OMTscore (C-index = 0.60, p = 0.02), which is comparable to the World Health Organization (WHO) grading (C-index = 0.64, p = 0.003) and binary grades (C-index = 0.65, p < 0.001). Notably, nuclear analyses revealed significant associations with peri-epithelial and intraepithelial lymphocytes in predictive patches of transforming cases (p < 0.001). This study represents the first fully automated and explainable algorithm for predicting OED transformation, offering a promising advancement in the clinical management of oral squamous cell carcinoma (OSCC) and enhancing early detection efforts critical for improving patient outcomes.
Methods
The “Methods” section outlines the experimental and analytical approaches employed in the study. It details the design of the experiments, including the selection of participants, materials used, and the specific procedures followed to ensure reliability and validity of the results. Statistical analyses were conducted using appropriate software to evaluate the data, with significance levels set at p < 0.05. Additionally, the section describes any mathematical models or equations applied to interpret the findings, ensuring that the methodologies align with the research objectives. The rigor of the methods is emphasized to support the credibility of the conclusions drawn from the study.
Results
In this study, we developed a multi-step pipeline to predict the OED risk score (OMTscore) using a deep learning approach. Initially, we trained an in-house deep learning model to segment intra-epithelial layers and nuclei. This model was subsequently applied to generate segmentations across all slides in our cohorts.
After segmentation, each slide was tessellated into smaller tiles, and tile-level morphological features were extracted based on the nuclear segmentations within the epithelium. These features were then utilized in a multilayer perceptron (MLP) to predict the likelihood of transformation to malignancy, thereby calculating the OMTscore. This approach highlights the integration of advanced image analysis techniques with predictive modeling to assess cancer risk.
Discussion
In this study, we introduced the OMTscore, a novel risk assessment tool designed to predict the transformation of oral epithelial dysplasia (OED) into malignancy. The model pipeline commenced with the simultaneous segmentation of nuclear instances and intra-epithelial layers using the HoVer-Net+ model on H&E-stained whole slide images (WSIs). The segmentation performance was validated on an internal Sheffield dataset, yielding satisfactory results that justified its application to both internal and external cohorts. Following segmentation, we generated patch-level morphological and spatial features, which were then utilized in a multi-layer perceptron (MLP) to produce the OMTscore. The model demonstrated competitive performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.77 on internal validation, surpassing traditional grading systems.
Survival analyses further confirmed the prognostic utility of the OMTscore, which exhibited a higher concordance index (C-index) compared to both WHO and binary grading systems, particularly in distinguishing low- from high-risk cases. Notably, the OMTscore’s performance was consistent across external validation cohorts, although some variability was observed, likely due to differences in sample characteristics and the generalizability of the segmentation model. Feature analyses revealed that transforming cases exhibited higher counts of specific nuclear types and a distinct cellular composition, suggesting that immune cell infiltration may play a role in malignant transformation. Overall, our findings underscore the OMTscore’s potential as a robust, interpretable tool for early detection of high-risk lesions, with implications for improving patient management and treatment strategies in OED. Future work should focus on enhancing model generalizability and validating the OMTscore in larger, multi-centric datasets.
