نموذج أساسي لتشخيص الأمراض النسيجية الحاسوبية ذات الجودة السريرية واكتشاف السرطانات النادرة
A foundation model for clinical-grade computational pathology and rare cancers detection

المجلة: Nature Medicine، المجلد: 30، العدد: 10
DOI: https://doi.org/10.1038/s41591-024-03141-0
PMID: https://pubmed.ncbi.nlm.nih.gov/39039250
تاريخ النشر: 2024-07-22
المؤلف: Eugene Vorontsov وآخرون
الموضوع الرئيسي: الذكاء الاصطناعي في اكتشاف السرطان

نظرة عامة

يتناول القسم التقدم في علم الأمراض الحاسوبي الذي يسهل استخدام مجموعات بيانات أكبر بكثير من تلك المستخدمة تاريخياً. يبرز دور خوارزميات التعلم الذاتي، التي تسمح بتدريب النماذج بشكل فعال دون الحاجة إلى بيانات موسومة بشكل واسع. مساهم رئيسي في تحسين أداء نماذج الرؤية الحاسوبية في هذا المجال هو تطوير الشبكات العصبية العميقة على نطاق واسع، المعروفة بالنماذج الأساسية. تستفيد هذه النماذج من كميات هائلة من البيانات لتحسين دقتها وكفاءتها في تطبيقات مختلفة ضمن علم الأمراض الحاسوبي.

الطرق

يستعرض قسم “الطرق” في ورقة البحث التصميم التجريبي والتقنيات التحليلية المستخدمة للتحقيق في أسئلة البحث. استخدمت الدراسة نهجاً كمياً، يتضمن تحليلات إحصائية لتقييم العلاقات بين المتغيرات. شملت جمع البيانات طريقة أخذ عينات منهجية، لضمان عينة تمثيلية من السكان قيد الدراسة.

استخدم الباحثون نماذج رياضية متنوعة لتحليل البيانات، بما في ذلك تحليل الانحدار لتحديد قوة وأهمية الارتباطات. بالإضافة إلى ذلك، استخدمت الدراسة أدوات برمجية لتصور البيانات، مما يسهل تفسير النتائج. تم تصميم المنهجية لضمان إمكانية إعادة الإنتاج والموثوقية، مع الالتزام بالإرشادات الأخلاقية المعمول بها طوال عملية البحث.

النتائج

قيمت نتائج الدراسة تضمينات نموذج فيرشو عبر فئتين رئيسيتين في علم الأمراض الحاسوبي على مستوى الشرائح: الكشف عن السرطان الشامل وتوقع العلامات الحيوية. بالنسبة للكشف عن السرطان الشامل، أشارت النتائج إلى أن نموذج فيرشو يسهل التعرف الفعال على أنواع السرطان المختلفة، مما يشير إلى إمكانيته لتحقيق أداء على مستوى العيادات. في سياق توقع العلامات الحيوية، أظهرت الأبحاث أن النموذج يمكنه اكتشاف العلامات الحيوية بدقة في التصوير الروتيني، مما يقلل من الحاجة إلى اختبارات إضافية.

لدعم هذه النتائج، استخدم الباحثون نموذج تجميع ضعيف الإشراف يجمع تضمينات البلاط لتوليد توقعات على مستوى الشرائح. بالإضافة إلى ذلك، تم إجراء سلسلة من معايير الاستكشاف الخطي على مستوى البلاط لتقييم قابلية تعميم التضمينات على البلاطات النسيجية الفردية. أكدت النتائج من هذه المعايير، إلى جانب التحليلات النوعية، على قوة وملاءمة نموذج فيرشو في مهام علم الأمراض الحاسوبي.

المناقشة

تسلط الأبحاث الضوء على تطوير نموذج للكشف عن السرطان الشامل باستخدام تضمينات فيرشو، والذي يظهر أداءً متفوقاً في التعرف على أنواع السرطان المختلفة، بما في ذلك السرطانات النادرة. تم تقييم النموذج باستخدام شرائح من مركز ميموريال سلوان كيترينغ للسرطان (MSKCC) ومؤسسات خارجية، محققاً منطقة تحت المنحنى (AUC) إجمالية تبلغ 0.950، مع تفوق تضمينات فيرشو على نماذج أخرى مثل UNI وPhikon وCTransPath عبر جميع أنواع السرطان. ومن الجدير بالذكر أن النموذج حافظ على AUC مرتفع يبلغ 0.937 للسرطانات النادرة، مما يبرز قوته وقدراته على التعميم، حتى عند اختباره على بيانات غير مدرجة في مجموعة التدريب.

في تحليل مقارن مع نماذج تجارية متخصصة، أدت نموذج الكشف عن السرطان الشامل القائم على فيرشو أداءً مماثلاً للنماذج المدربة خصيصاً لسرطانات البروستاتا والثدي، بينما تفوق في اكتشاف متغيرات السرطان النادرة. على الرغم من تدريبه على عدد أقل من العينات النسيجية المحددة، حقق نموذج السرطان الشامل AUCs تبلغ 0.980 و0.985 و0.971 لسرطانات البروستاتا والثدي والعقد اللمفاوية، على التوالي. تشير النتائج إلى أن تضمينات فيرشو لا تعزز فقط دقة الكشف عن السرطان ولكن أيضاً تسهل توقع العلامات الحيوية من الصور الملونة القياسية H&E، مما قد يقلل من الحاجة إلى طرق اختبار إضافية. بشكل عام، تؤكد الدراسة على وعد النماذج الأساسية في تحسين الكشف عن السرطان والتشخيص عبر أنواع الأنسجة المتنوعة.

Journal: Nature Medicine, Volume: 30, Issue: 10
DOI: https://doi.org/10.1038/s41591-024-03141-0
PMID: https://pubmed.ncbi.nlm.nih.gov/39039250
Publication Date: 2024-07-22
Author(s): Eugene Vorontsov et al.
Primary Topic: AI in cancer detection

Overview

The section discusses the advancements in computational pathology facilitated by the use of significantly larger datasets than those historically employed. It highlights the role of self-supervised learning algorithms, which allow for effective model training without the need for extensive labeled data. A key contributor to the enhanced performance of computer vision models in this field is the development of large-scale deep neural networks, known as foundation models. These models leverage vast amounts of data to improve their accuracy and efficiency in various applications within computational pathology.

Methods

The “Methods” section of the research paper outlines the experimental design and analytical techniques employed to investigate the research questions. The study utilized a quantitative approach, incorporating statistical analyses to assess the relationships between variables. Data collection involved a systematic sampling method, ensuring a representative sample of the population under study.

The researchers employed various mathematical models to analyze the data, including regression analysis to determine the strength and significance of correlations. Additionally, the study utilized software tools for data visualization, facilitating the interpretation of results. The methodology was designed to ensure reproducibility and reliability, adhering to established ethical guidelines throughout the research process.

Results

The results of the study evaluated the Virchow model embeddings across two primary categories in slide-level computational pathology: pan-cancer detection and biomarker prediction. For pan-cancer detection, the findings indicated that the Virchow model facilitates effective identification of various cancer types, suggesting its potential for achieving clinical-grade performance. In the context of biomarker prediction, the research demonstrated that the model can accurately detect biomarkers in routine imaging, thereby reducing the need for additional testing.

To support these findings, the researchers employed a weakly supervised aggregator model that aggregates tile embeddings to generate slide-level predictions. Additionally, a series of tile-level linear probing benchmarks were conducted to assess the generalizability of the embeddings on individual tissue tiles. The results from these benchmarks, along with qualitative analyses, further confirmed the robustness and applicability of the Virchow model in computational pathology tasks.

Discussion

The research highlights the development of a pan-cancer detection model utilizing Virchow embeddings, which demonstrates superior performance in identifying various cancer types, including rare cancers. The model was evaluated using slides from the Memorial Sloan Kettering Cancer Center (MSKCC) and external institutions, achieving an overall area under the curve (AUC) of 0.950, with Virchow embeddings outperforming other models such as UNI, Phikon, and CTransPath across all cancer types. Notably, the model maintained a high AUC of 0.937 for rare cancers, showcasing its robustness and generalization capabilities, even when tested on data not included in the training set.

In a comparative analysis with specialist commercial models, the Virchow-based pan-cancer detection model performed comparably to models specifically trained for prostate and breast cancers, while excelling in detecting rare cancer variants. Despite being trained on fewer tissue-specific specimens, the pan-cancer model achieved AUCs of 0.980, 0.985, and 0.971 for prostate, breast, and lymph node cancers, respectively. The findings suggest that the Virchow embeddings not only enhance cancer detection accuracy but also facilitate biomarker predictions from standard H&E stained images, potentially reducing the need for additional testing methods. Overall, the study underscores the promise of foundation models in improving cancer detection and diagnosis across diverse tissue types.