Thesis (M.Sc.)-Cairo University, 2024.

Bibliography: pages 85-94.

Cancer is the second-leading cause of death globally, only behind heart diseases.
Breast cancer is one of the most common and aggressive types of cancer, but it is highly
treatable if caught early. To achieve early diagnosis with effective results, fine-needle
aspiration is typically carried out by pathologists, who collect tissue samples, visually
examine geometric and textural features, and thus identify an examined lesion as either
benign or malignant. However, this manual lesion classification process is highly costly
and time-consuming, and its outcomes also vary depending on the pathologist’s
expertise. To alleviate this burden, numerous methods have been proposed for automatic
computer-aided diagnosis (CAD) of breast cancer with different types of features, feature
selection methods, and classification methods. In this thesis, a new approach for breast
cancer detection is introduced where integrated ranking techniques are employed to
aggregate the feature ranks obtained from ensembles of different feature selection
methods. In particular, six state-of-the-art feature selection methods (along with four data
scaling methods, ten types of classifiers, and wide ranges of classifier hyperparameters)
were initially explored. The performance under each classifier configuration was
evaluated based on ten different metrics. Then, as each of the six feature selection
methods returned a different ranking of the computed features, an effort was made to
combine the rankings within each ensemble of methods through five integrated ranking
schemes, namely, the Borda, Copeland, grade average, cross-entropy Monte Carlo, and
genetic algorithms. Compared to the approaches based on individual feature selection
methods, remarkable improvements in classification performance were demonstrated by
our integrated ranking approach for two commonly used breast cancer datasets.
Experiments show that the proposed model achieved the best accuracy of 100% for the
WDBC dataset and 100%, 99.75%, and 99.68% for the WBCD dataset.

سرطان الثدي هو أحد أكثر أنواع السرطان شيوعاً وخطورة وسرعة في الانتشار خلال جسم المريض، إلا أنّه يمكن علاجه بشكل كبير إذا تم تشخصيه مبكراً. قدمنا تشخيصًا يعتمد على الذكاء الاصطناعي للكشف عن سرطان الثدي بناءً على تجميع الرتب ضمن مجموعات طرق اختيار المعالم. استكشفنا في البداية ست طرق حديثة لاختيار المعالم .تم تقييم أداء كل مُصنِف على أساس عشرة مقاييس مختلفة. نظرًا لأن كل طريقة من الطرق الست لاختيار المعالم أنتجت ترتيبًا مختلفًا للمعالم، سعينا إلى الجمع بين الترتيبات المختلفة من خلال ست طرق لتجميع الرتب. أظهر نهجنا المتكامل في تجميع الرتب تحسيناتٍ ملحوظةً في معايير تقييم أداء المُصنِفات لمجموعتين شائعتي الاستخدام من بيانات سرطان الثدي، حيث وصلت دقة النموذج المقترح على كلٍ من المجموعتين إلى 100٪,99.75%, 99.68%.

Issued also as CD

Text in English and abstract in Arabic & English.