Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions

Deep learning has revolutionized medical image analysis, particularly in the domain of breast cancer detection. Despite notable progress, further optimization of neural network architectures and activation functions remains critical for enhancing classification accuracy and model generalization. Thi...

Full description

Saved in:
Bibliographic Details
Main Authors: Khaddouj Taifi, Yassine Sabbar, Hanin Ardah, Abdel-Haleem Abdel-Aty
Format: Article
Language:English
Published: IEEE 2025-01-01
Series:IEEE Access
Subjects:
Online Access:https://ieeexplore.ieee.org/document/11106463/
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1849389757294968832
author Khaddouj Taifi
Yassine Sabbar
Hanin Ardah
Abdel-Haleem Abdel-Aty
author_facet Khaddouj Taifi
Yassine Sabbar
Hanin Ardah
Abdel-Haleem Abdel-Aty
author_sort Khaddouj Taifi
collection DOAJ
description Deep learning has revolutionized medical image analysis, particularly in the domain of breast cancer detection. Despite notable progress, further optimization of neural network architectures and activation functions remains critical for enhancing classification accuracy and model generalization. This study explores the effect of architectural refinements and activation function choices on the performance of three widely adopted convolutional neural networks: DenseNet-121, DenseNet-201, and MobileNetV2. Specifically, we introduce a modified framework that incorporates selective freezing of early convolutional layers and replaces the conventional ReLU activation function with the smoother GELU activation in the fully connected layers, aiming to improve feature representation and classification robustness. The proposed enhancements are rigorously evaluated across three balanced and widely used benchmark mammography datasets: MIAS, INbreast, and DDSM, using standard performance metrics including Precision, Recall, F1-score, and Accuracy. Experimental findings show that the modified models consistently surpass their original counterparts. Notably, the enhanced DenseNet-201 achieves the highest accuracy of 99.6% on the DDSM dataset. Similarly, the modified DenseNet-121 attains 97% accuracy on MIAS and 98% on INbreast, while the improved DenseNet-201 reaches 98% on both MIAS and INbreast. Furthermore, the lightweight yet optimized MobileNetV2 achieves impressive results, with 99.4% accuracy on DDSM and 99% on INbreast. To validate the statistical significance of these performance gains, paired t-tests were conducted, confirming that the improvements are not only consistent but also statistically meaningful. Our results emphasize the effectiveness of architectural tuning and activation function refinement in advancing deep learning-based breast cancer classification.
format Article
id doaj-art-76bf681fa3f8433da3834994eb38ac4d
institution Kabale University
issn 2169-3536
language English
publishDate 2025-01-01
publisher IEEE
record_format Article
series IEEE Access
spelling doaj-art-76bf681fa3f8433da3834994eb38ac4d2025-08-20T03:41:52ZengIEEEIEEE Access2169-35362025-01-011313785113786810.1109/ACCESS.2025.359482711106463Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation FunctionsKhaddouj Taifi0https://orcid.org/0009-0001-0221-9714Yassine Sabbar1https://orcid.org/0000-0002-1127-4395Hanin Ardah2https://orcid.org/0009-0007-7123-9616Abdel-Haleem Abdel-Aty3https://orcid.org/0000-0002-6763-2569Department of Informatics, FST Errachidia, IPDAC Team, Moulay Ismail University of Meknes, Errachidia, MoroccoDepartment of Mathematics, FST Errachidia, IMIA Laboratory, T-IDMS, Moulay Ismail University of Meknes, Errachidia, MoroccoDepartment of Computer Sciences, College of Computer and Information Sciences, Princess Nourah bint Abdulrahman University, P.O. Box 84428, Riyadh, Saudi ArabiaDepartment of Physics, College of Sciences, University of Bisha, Bisha, Saudi ArabiaDeep learning has revolutionized medical image analysis, particularly in the domain of breast cancer detection. Despite notable progress, further optimization of neural network architectures and activation functions remains critical for enhancing classification accuracy and model generalization. This study explores the effect of architectural refinements and activation function choices on the performance of three widely adopted convolutional neural networks: DenseNet-121, DenseNet-201, and MobileNetV2. Specifically, we introduce a modified framework that incorporates selective freezing of early convolutional layers and replaces the conventional ReLU activation function with the smoother GELU activation in the fully connected layers, aiming to improve feature representation and classification robustness. The proposed enhancements are rigorously evaluated across three balanced and widely used benchmark mammography datasets: MIAS, INbreast, and DDSM, using standard performance metrics including Precision, Recall, F1-score, and Accuracy. Experimental findings show that the modified models consistently surpass their original counterparts. Notably, the enhanced DenseNet-201 achieves the highest accuracy of 99.6% on the DDSM dataset. Similarly, the modified DenseNet-121 attains 97% accuracy on MIAS and 98% on INbreast, while the improved DenseNet-201 reaches 98% on both MIAS and INbreast. Furthermore, the lightweight yet optimized MobileNetV2 achieves impressive results, with 99.4% accuracy on DDSM and 99% on INbreast. To validate the statistical significance of these performance gains, paired t-tests were conducted, confirming that the improvements are not only consistent but also statistically meaningful. Our results emphasize the effectiveness of architectural tuning and activation function refinement in advancing deep learning-based breast cancer classification.https://ieeexplore.ieee.org/document/11106463/Deep learningbreast cancer detectionDenseNet-121DenseNet-201MobileNetV2GELU
spellingShingle Khaddouj Taifi
Yassine Sabbar
Hanin Ardah
Abdel-Haleem Abdel-Aty
Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions
IEEE Access
Deep learning
breast cancer detection
DenseNet-121
DenseNet-201
MobileNetV2
GELU
title Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions
title_full Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions
title_fullStr Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions
title_full_unstemmed Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions
title_short Enhanced DL-Based Breast Cancer Diagnosis and Classification Using Modified DenseNet-121, DenseNet-201, and MobileNetV2: Optimized Architectures and Refined Activation Functions
title_sort enhanced dl based breast cancer diagnosis and classification using modified densenet 121 densenet 201 and mobilenetv2 optimized architectures and refined activation functions
topic Deep learning
breast cancer detection
DenseNet-121
DenseNet-201
MobileNetV2
GELU
url https://ieeexplore.ieee.org/document/11106463/
work_keys_str_mv AT khaddoujtaifi enhanceddlbasedbreastcancerdiagnosisandclassificationusingmodifieddensenet121densenet201andmobilenetv2optimizedarchitecturesandrefinedactivationfunctions
AT yassinesabbar enhanceddlbasedbreastcancerdiagnosisandclassificationusingmodifieddensenet121densenet201andmobilenetv2optimizedarchitecturesandrefinedactivationfunctions
AT haninardah enhanceddlbasedbreastcancerdiagnosisandclassificationusingmodifieddensenet121densenet201andmobilenetv2optimizedarchitecturesandrefinedactivationfunctions
AT abdelhaleemabdelaty enhanceddlbasedbreastcancerdiagnosisandclassificationusingmodifieddensenet121densenet201andmobilenetv2optimizedarchitecturesandrefinedactivationfunctions