Cervical cancer is the second most common cancer among women and a leading cause of mortality. Many attempts have been made to develop an effective Computer-Aided Diagnosis (CAD) system, however, their performance remains limited. Using pretrained ResNet-50/101/152, we propose a novel CAD system that significantly outperforms prior approaches. Our model has three key components. First, as common in a lot of cancer classification systems, we extract the detailed features (color, edges, and texture) from early convolution blocks and the abstract features (shapes and objects) from later blocks, as both are equally important. Second, a non-parametric 3D attention module is embedded within each block for feature enhancement. Embedding attention within convolution blocks is also common in cancer classification, however, the way we integrate it with the dual-level feature extraction technique is new. Third, we design a theoretically motivated adaptive pooling strategy for feature selection that applies Global Max Pooling to detailed features and Global Average Pooling to abstract features. Typically, these pooling layers are interspersed between convolution blocks, as common in different cancer systems. However, the way we adapt these pooling strategy for dual-level feature extraction technique is also new. These components form our proposed Block-Fused Attention-Driven Adaptively-Pooled ResNet (BF-AD-AP-ResNet) model. To further strengthen learning, we introduce a Tri-Stream model, as common in the cancer literature, which unifies the enhanced features from three BF-AD-AP-ResNets. An SVM classifier is employed for final classification. We evaluate our models on two public datasets, IARC and AnnoCerv. On IARC, the base ResNets achieve an average performance of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$83.61\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>83.61</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, while our model achieves an excellent performance of <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$ {\textbf {91.24\%}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn mathvariant="bold">91.24</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>. On AnnoCerv, the base ResNets reach to <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$78.75\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>78.75</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, and our model improves this significantly, reaching <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$ {\textbf {86.35\%}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn mathvariant="bold">86.35</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>. Our approach outperforms the best existing method on IARC by an average of <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$ {\textbf {7\%}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn mathvariant="bold">7</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>. For AnnoCerv, no prior competitive works are available. We conduct ablation studies to justify the inclusion of each component. Additionally, we introduce a SHAP+LIME explainability method, accurately identifying the cancerous region in <InlineEquation ID="IEq6"> <EquationSource Format="TEX">${\textbf {97\%}}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn mathvariant="bold">97</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> of cases. Although, SHAP and LIME have been independently applied in different cancer studies, however, the manner we integrate them is new as well.

Block-fused attention-driven adaptively-pooled ResNet model for improved cervical cancer classification

期刊论文
发表日期：

2026年5月27日

查看全文

Saurabh Saini,
Kapil Ahuja,
Akshat S. Chauhan

摘要

Cervical cancer is the second most common cancer among women and a leading cause of mortality. Many attempts have been made to develop an effective Computer-Aided Diagnosis (CAD) system, however, their performance remains limited. Using pretrained ResNet-50/101/152, we propose a novel CAD system that significantly outperforms prior approaches. Our model has three key components. First, as common in a lot of cancer classification systems, we extract the detailed features (color, edges, and texture) from early convolution blocks and the abstract features (shapes and objects) from later blocks, as both are equally important. Second, a non-parametric 3D attention module is embedded within each block for feature enhancement. Embedding attention within convolution blocks is also common in cancer classification, however, the way we integrate it with the dual-level feature extraction technique is new. Third, we design a theoretically motivated adaptive pooling strategy for feature selection that applies Global Max Pooling to detailed features and Global Average Pooling to abstract features. Typically, these pooling layers are interspersed between convolution blocks, as common in different cancer systems. However, the way we adapt these pooling strategy for dual-level feature extraction technique is also new. These components form our proposed Block-Fused Attention-Driven Adaptively-Pooled ResNet (BF-AD-AP-ResNet) model. To further strengthen learning, we introduce a Tri-Stream model, as common in the cancer literature, which unifies the enhanced features from three BF-AD-AP-ResNets. An SVM classifier is employed for final classification. We evaluate our models on two public datasets, IARC and AnnoCerv. On IARC, the base ResNets achieve an average performance of $83.61\%$ $83.61 %$ , while our model achieves an excellent performance of $ {\textbf {91.24\%}}$ $91.24 %$ . On AnnoCerv, the base ResNets reach to $78.75\%$ $78.75 %$ , and our model improves this significantly, reaching $ {\textbf {86.35\%}}$ $86.35 %$ . Our approach outperforms the best existing method on IARC by an average of $ {\textbf {7\%}}$ $7 %$ . For AnnoCerv, no prior competitive works are available. We conduct ablation studies to justify the inclusion of each component. Additionally, we introduce a SHAP+LIME explainability method, accurately identifying the cancerous region in ${\textbf {97\%}}$ $97 %$ of cases. Although, SHAP and LIME have been independently applied in different cancer studies, however, the manner we integrate them is new as well.