Early breast cancer detection in CT scans using convolutional neural bidirectional feature pyramid network

Introduction

Despite various approaches for detection and prevention, the incidence of human illnesses has progressively risen due to individual or environmental factors. Among these illnesses, cancer is the most prevalent global affliction, characterized by abnormal cell growth that impacts different body areas. Cancer manifests in various forms, including breast, prostate, lung, skin, and pancreatic cancer, and remains a leading cause of death worldwide (Rajinikanth et al., 2021). Early detection is crucial, as undiagnosed breast cancer can be fatal. Advances in early diagnosis and treatment have significantly improved survival rates. Various imaging techniques are employed for breast cancer screening, including X-ray radiography, computed tomography (CT) scans, ultrasonography, magnetic resonance imaging (MRI), and thermal imaging, each offering unique advantages in detecting abnormalities.

Each modality employs unique methods and tools, yielding results influenced by various factors, making it advisable to combine multiple methods for result validation. Mammography, the most prevalent screening procedure for breast cancer, is often considered the benchmark, but it poses considerable risks to patients. Consequently, thermal imaging has garnered attention due to its lower risk profile than mammography (Singh & Singh, 2020). Despite not yet being the primary early breast cancer detection method, digital infrared thermal imaging has shown promise. While relying on skin lesions and temperature, it can complement mammography when used in conjunction, enabling radiologists to assess and diagnose breast cancer (Parisky et al., 2000) more accurately.

Artificial neural networks (ANNs) employ mathematical models and emulate the structure of mammalian neural systems on a smaller scale (LeCun, Bengio & Hinton, 2015; Yamashita et al., 2018). Whether object detection, image classification, or image segmentation, convolutional neural networks (CNNs) have demonstrated superior capabilities in automatically learning and extracting relevant image features (Bhojanapalli et al., 2021).

Transformer-based architectures have recently shown great potential as breast cancer detection tools, significantly improving traditional algorithms such as CNNs. Unlike CNNs, which are based on local feature extraction using fixed spatial extent of the kernel, transformers use self-attention mechanisms to capture arbitrary-length context across the whole image. This capability enables transformer models like Vision Transformers (ViTs) and Swin Transformers to examine complex patterns in the medical imaging data such as mammograms, and CT scans. Recent studies have shown that ViTs outperform CNNs in classifying breast cancer lesions since they can develop better feature extraction and understanding of spatial relation across various locations in an image (Dosovitskiy et al., 2020). Swin Transformer, which proposes hierarchical feature learning through shifted windows, also boosts tumor segmentation accuracy, assisting radiologists in obtaining accurate and early diagnoses (Liu et al., 2021). Hybrid architectures, combining CNNs and transformers, have improved breast cancer diagnosis even more by aligning the regional feature extraction performed by CNNs with the contextual features offered by transformers (Liu et al., 2021).

Traditional imaging methods, such as mammography, are widely used for breast cancer detection but suffer from drawbacks like radiation exposure, high false positive/negative rates, and challenges in detecting tumors in dense breast tissue. While CT scans offer high-resolution imaging, extracting meaningful patterns requires advanced computational techniques. Though effective in feature extraction, conventional CNN-based approaches struggle with detecting tumors of varying sizes due to their fixed receptive fields and inefficient multi-scale feature utilization, making them less robust for complex medical imaging tasks. To address this limitation, our study integrates bidirectional feature pyramid network (BiFPN) (Tan & Le, 2019; Redmon & Farhadi, 2018), which enhances feature fusion by leveraging multi-scale representations, enabling better tumor localization and improving detection accuracy. Unlike standard CNNs, BiFPN refines features bidirectionally, capturing both fine-grained details and high-level contextual information, thus providing a more reliable approach for breast cancer detection in CT scans.

Related work

In recent years, many machine-learning techniques have been applied to analyze breast cancer imaging. One innovative approach combined convolutional neural networks and long short-term memory networks to classify over 7,000 histological slides. After extracting features, the proposed model fed them into softmax and support vector machine classifiers achieving an impressive 96% precision on low magnification images and 91% highest accuracy on high magnification, consistent with advances. Another study (Rashed & El Seoud, 2019) introduced a novel U-net inspired convolutional neural network designed for mammography datasets focusing on masses and calcifications. The network trained on nearly 700 mass images and over 600 calcification images, reserving over 200 and 150 respectively for testing, achieving a remarkable 94.31% accuracy. An additional innovative model, IRRCNN (Arslan, Yaşar & Çolak, 2019), was a convolutional neural network framework to classify breast cancer histopathology using the same dataset, achieving 91.4% accuracy during testing and demonstrating effectiveness. Furthermore, BC-DROID enables fully automated detection and classification in a single step using convolutional neural networks. Trained on over 10,000 complete mammograms, BC-DROID achieved 90% detection accuracy, 93.5% classification accuracy and 92.315% area under the curve.

A hybrid deep learning model named CRNN, described in Patil & Biradar (2021), was used to detect breast cancer from mammographic images. The process was to build a gray level cooccurrence matrix and a gray level run length matrix as feature inputs. Using a series of SELECT && operations, the diagnostic results derived by the classifier proved to be accurate in 90.59% of cases. Compared with traditional models, this accuracy is superior by far. With the rapid development of infrared camera technology, research on diagnosing and classifying breast cancer based on thermal images has also made significant progress. Many studies have used machine learning (ML) and deep learning (DL) algorithms to analyze this form of images.

A comprehensive analysis of 1,052 thermogram images was conducted at The Federal University of Pernambuco University Hospital in 2015 (Santana et al., 2018). Bayes network, naïve Bayes, support vector machine (SVM), J48 decision tree algorithm, random forest (RF), multi-layer perceptron, random forest tree, among these classifiers, MLP is a good prospect with accuracy rate of 73.38%, kappa index value reaches 0.6007, sensitivity 78% and specificity 88%. Furthermore, after reflecting changes using 10-fold cross-validation, there was a better kappa index value 0.6402 now at an improved accuracy rate 76.01%. Overall, system effectiveness improved from 83% to 84%.

To enhance the data classification with ALR and KCC for DBT-TU-JU and DMR-IR series, classifier algorithms artificial neural network (ANN), SVM, k-nearest neighbor (KNN) and decision trees (DT) were improved with features SSigFS, FStat, and STex (Gogoi et al., 2017). The SVM-RBF and ANN models scored a better accuracy rate of 84.29% on DVDTU-JU series, While ANN and SVM-linus are the best classifiers in k-fold using DSIE-IR based and picked 87.50% accuracy rate. Deep learning (DL) techniques, especially convolutional neural networks (CNNs), have also been employed to study thermogram images. In Ekici & Jawzal (2020), a CNN was trained on histopathology images and fine-tuned by Bayes optimization. Authors collected 1,116 images from the DMI data set, the method achieved a remarkable minimum error rate of 98.05%, which outshone earlier reports using different feature sets and browse classification on the same dataset.

In another work (Mishra et al., 2020), employed a large DCNN designed especially for breast cancer diagnosis from thermography images. To concisely formalise the results, 680 images from the Visual lab-IR data set converted to grayscale images were effectively segmented and categorized. The forecasting accuracy rate is 95.8%, which is better than (Mambou et al., 2018). A study in 2012 (Mookiah, Acharya & Ng, 2012) used decision trees and fuzzy classifiers with an accuracy 92.30%.

Tello-Mijares, Woo & Flores (2019) investigated a combination of generalized vector field (GVF) breast segmentation with CNN for breast cancer classification (Nahid, Mehrabi & Kong, 2018). In this study, 63 breast images from DMR-IR dataset were used, which were classified as Normal and Abnormal. With 2-fold cross-validation, the model delivered outstanding results, with accuracy, sensitivity, and specificity, all at an extraordinary 100% level. Importantly, their performance was higher than that of tree random forest (TRF), multi-layer perceptron (MLP), and Bayes network classifiers. For Sanyal, Jethanandani & Sarkar (2020) an attention mechanism (AM) was used to perform the categorization of high-resolution histological images from breast tissue with (CNN + bidirectional long short-term memory (BLSTM)). The model’s training was conducted on the ICIAR 2018 database. The experimental findings unveiled that the model attained an accuracy of 85.50% for classifying images into four classes and an even more remarkable accuracy of 96.25% for two classes. A comparative analysis was performed with and without the AM and against state-of-the-art methods. Notably, the model consistently outperformed other approaches with and without integrating the attention mechanism. In Deng et al. (2020), an enhancement was made to CNNs by incorporating an innovative SE-attention mechanism, which proved highly effective in classifying a collection of 18,157 mammograms. A new benchmarking dataset was curated for this purpose, and the model demonstrated superior performance compared to other studies, achieving an accuracy rate of 92.17%. To enhance performance on the BreakHis dataset, Zhang et al. (2019) proposed a classification framework for breast cancer tissue images based on ResNet integrated with a convolutional block attention module (CBAM). For improving performance on the BreakHis dataset, Zhang et al. (2019) designed a classification framework using RESNET integrated with CBAM to classify images of breast cancer tissue. The model is then used to classify the 7,909 images with significant increases to an accuracy of 92.6%, sensitivity of 94.7%, specificity of 88.9%, F1-score of 94.1%, and an AUC of 91.8% for 200× magnified images. Additionally, the proposed model outperformed ResNet50 and ResNet-50 with CBAM in terms of accuracy by a large margin at both the patient and image levels. Franceschini et al. (2023) introduced a deep learning method for tumor detection in a microwave tomography framework, yielding promising results in identifying small tumor masses. Gengtian, Bing & Guoyou (2023) applied the EfficientNet architecture for detecting and classifying breast cancer in mammography images, highlighting the effectiveness of deep learning in medical imaging. Additionally, deep learning models like CNN and Vgg16 for evaluating histopathology images have demonstrated great potential in improving both the accuracy and speed of breast cancer detection.

The hybrid model developed by Ametefe et al. (2025) connected deep transfer learning methods with U-Net segmentation to find breast cancer through ultrasound reviews. A main drawback of this approach is that it exclusively uses ultrasound images, which prevents direct application across other imaging systems. Al-Duais et al. (2025) discussed the development of a comparative approach of classical machine learning technologies like SVM, KNN, and RF appropriate for early breast cancer detection. Fontes et al. (2025) presented a new 3D CNN method designed for luminal A breast cancer phenotyping on CT and MRI images. Their model greatly improved the classification accuracy compared to traditional 2D CNNs by considering spatial features. Marinakis et al. (2025) discussed the precision challenge in a study on 3D residual networks for segmenting pulmonary nodules in low-dose CT scans whereas, initially targeted to lung cancer, which is also acceptable for detecting breast lesions. Their data generation with the patch helped the model to be much more sensitive to small lesions, a key requirement for early screening for breast cancer (Marinakis et al., 2025). A multi-modal fusion system composed of mammography, MRI and CT scan has been proposed by Lew et al. (2025) with the intentions of increasing breast cancer detection accuracy. Wang et al. (2025) participated in the RSNA Breast Cancer Detection competition and applied deep learning model on CT data to predict mammographic breast density. Although not a primary detection task, accurate density assessment is a strong indicator of cancer risk and their model provided an important auxiliary tool for risk stratification (Wang et al., 2025). Sun et al. (2025) proposed DL models based on chest CT images to differentiate benign and malignant breast masses and predict axillary lymph node metastasis. Their strategy offered a non-invasive substitute for biopsy, which had substantial potential for clinical use but also stressed the significant need for large-scale validation (Sun et al., 2025). Tong et al. (2025) introduced the integration of machine learning with panoramic photoacoustic computed tomography (PACT) for breast lesion analysis. Although primarily experimental, this technique offered an innovative alternative to conventional CT and showed that DL could improve lesion boundary detection and characterization (Tong et al., 2025).

Breast cancer detection through deep learning has been thoroughly investigated, with various studies concentrating on CNN-based frameworks, feature extraction methods, and segmentation techniques. While prior research has achieved notable advancements, significant gaps still need to be addressed. Despite their robust feature extraction abilities, conventional CNN architectures frequently face challenges in detecting tumors across multiple scales, particularly in intricate medical imaging contexts where tumors exhibit considerable variation in size, shape, and contrast. The work conducted by Lin et al. (2017a) introduced feature pyramid networks (FPNs) to tackle the issue of multi-dimension feature learning; however, these models experience limitations due to unidirectional feature propagation, resulting in suboptimal feature fusion. Likewise, alternative methods, including U-Net and Mask R-CNN, have been utilized for medical image segmentation but no dynamic feature reweighting (He et al., 2017; Ronneberger, Fischer & Brox, 2015).

The proposed framework incorporated the BiFPN with CNNs, facilitating effective multi-scale feature fusion through iterative bidirectional information flow. In contrast to conventional FPNs, BiFPN enhances feature selection via dynamic weighted connections, guaranteeing that only the most pertinent features play a role in the final classification and segmentation processes (Tan, Pang & Le, 2020). This pivotal improvement boosts tumor localization, classification precision, and reduces false positives, which are frequent limitations of current CNN-based models. Mughal et al. (2018) proposed a novel classification system to reduce breast tumor-related mortality among women. Their approach demonstrated enhanced diagnostic reliability using microscopy imaging techniques (Mughal et al., 2018). Elkorany et al. (2022) utilized SVM optimized by whale optimization algorithm (WOA) and dragonfly algorithm (DA) for breast cancer diagnosis. The dual-optimization technique enhanced classification performance significantly over standard SVM (Elkorany et al., 2022). Marie-Sainte et al. (2019) developed an improved prediction model for breast cancer diagnosis, survivability, and recurrence. The strategy combined data mining techniques with machine learning to forecast patient outcomes. Results showed that the model outperformed conventional methods in predictive accuracy and decision support (Marie-Sainte et al., 2019). Table 1 describes the literature review with drawbacks of the existing approaches.

Proposed framework

We developed a novel hybrid framework called the convolutional neural bidirectional feature pyramid network (CNBiFPN), which incorporates four key stages, as shown in Fig. 1. First, the data collection module describes the accumulation of inputs. The data is collected from Ayub Teaching Hospital Complex Abbottabad, Pakistan through CT Scans images. We have also received the informed consent from participants in our experimental study. The Institutional Review Board approved our human study with reference number ATH/1101-H/2022. Secondly, the data preprocessing phase carried out crucial transformations on the images using techniques designed to adequately prepare them for subsequent steps. Next, the feature extraction and classification module leveraged a convolutional neural network to derive representations from the images and then categorize them as either diseased or healthy. Finally, the image segmentation component employed a bidirectional feature pyramid network to both demarcate and localize the affected regions within the images with precision. Our framework holds promise as a powerful tool for the careful and targeted analysis of medical images.

Experimental setup and results

We have conducted experiments on actual dataset collected from Radiology Department, Ayub Teaching Hospital Abbottabad, Pakistan. We collected about 500 CT scan images related to normal patients and 500 related to abnormal (diseased) patients, and supplemented them with additional anonymized scans provided by a collaborating hospital and verified by expert radiologists. The dataset includes normal and diseased images, labelled with 0 for normal and 1 for diseased. The dataset was split into 70% training, and 30% testing sets using a stratified split to maintain class balance. The dataset and software code is available at https://github.com/aahmedqau/Breast-Cancer-Detection-using-CT-Scans.

Results and analysis

This section presents the evaluation of results using the proposed model and compares them against baseline approaches. We randomly selected 70% of the instances to train the machine learning models. The remaining 30% of the isolated dataset was used to test how well the model performs on new data.

Image classification

Figure 2 shows the classification of breast cancer CT scan images into diseased and non-diseased images.

Image segmentation

When processing images through machine learning, it is common for a medical dataset to be segmented from the original medical image, using a suitable medical image segmentation method to obtain the specific region of interest (Hoffman & Jain, 1987). It separates regions based on their similarity or differences in terms of image intensity.

Model training

Model training is performed as follows.

Training with batch 0

Figure 3 illustrates the training with the initial batch 0, utilizing 16 breast cancer-infected images from patients. The enclosed area denotes the actual infected region diagnosed by an expert, whereas the square area indicates the predicted region identified by the proposed model.

Training with batch 1

Figure 4 illustrates the training process using 16 breast cancer-infected images, processed in batch 1. Encircled region highlights the original diseased area diagnosed by an expert, while the square region indicates the predicted area.

Training with batch 2

Figure 5 illustrates training process using 16 images of patients with infections (Batch 2). The highlighted area indicates the region initially diagnosed as infected by an expert, while the square region represents the predicted area identified by the proposed model.

Model validation

We labeled each image based on the infected region and trained the data accordingly. Each image’s infected region was predicted using the proposed model. Figure 6 shows a red circle marker indicating the actual infected region in the CT scan of a breast cancer patient, and a red square marker representing the predicted region of the cancerous area detected by the proposed model.

Performance comparison using evaluation metrics

To create a precision-recall curve for the classification of breast cancer, precision and recall metrics are computed at various classification thresholds, corresponding to the probabilities used for determining whether a sample is labeled as “cancerous” or “non-cancerous.” An assessment of the proposed model’s performance using the precision-recall metric is depicted in Fig. 7. The graph represents how precision and recall values change relative to the threshold. Ensuring high precision is crucial to prevent false positives, which could lead to unnecessary treatment and patient distress. Similarly, achieving high recall is essential to prevent false negatives, which could result in missed diagnoses and poorer outcomes. Figure 8 exhibits the precision-recall curve during validation, along with MAP @0.5.

The proposed CNBiFPN framework demonstrated strong diagnostic performance, achieving a sensitivity of 94.84% and a specificity of 97.38%. The high sensitivity indicates that the model is highly capable of correctly identifying patients with breast cancer from CT scans, thereby minimizing missed diagnoses and reducing the risk of untreated disease progression. Meanwhile, the high specificity reflects the model’s effectiveness in accurately classifying healthy cases, which helps lower the incidence of false positives.

In Fig. 9, the training and validation loss curves show a consistent and stable decline across 100 epochs, indicating effective learning without overfitting. The minimal gap between the training and validation losses suggests strong generalization capabilities. The final loss values (~0.03 for training and ~0.034 for validation) demonstrate that the model successfully minimized prediction errors, achieving high reliability on both seen and unseen data.

The F1 score serves as an evaluation metric for gauging the performance of a model. It strikes a balance between precision and recall, with precision assessing the model’s accuracy in predicting positives and recall, evaluating its capability to identify all positive instances, such as breast cancer cases. In Fig. 10, the F1 confidence-curve illustrates F1 measure values ranging from 0.00066546 to 0.94485. Higher F1 scores signify a more optimal equilibrium between precision and recall, implying a more dependable model for early detection of breast cancer.

Statistical analysis

We have validated the results of proposed framework using confidence intervals and Cohen’s Kappa measure.

Validation of results using confidence intervals

We obtained accuracy of proposed framework is about 0.9611 and the sampling is 70% training and 30% testing of 1,000 images. So, confidence intervals for accuracy can be computed as

(15) $$CI=\hat{p}\pm Z\times \sqrt{{\displaystyle \frac{\hat{p}\left(1-\hat{p}\right)}{n}}}$$where $\hat{p}$ is the observed accuracy that is 0.9611, $n$ is number of samples that is 300 and $Z$ = 1.96 for 95% confidence.

Since, the observed accuracy is 96.11% and 95% confidence interval for accuracy ≈ [93.92%, 98.30%]. Hence, we are 95% confident that the true accuracy of the model lies between 93.92% and 98.30% on unseen CT scan images for breast cancer detection.

Validation of results using Cohen’s Kappa measure

We also performed validation between ground truth labels marked by radiologist in the form of circle and the locations predicted by proposed model in the form of square using Cohen’s Kappa metric:

(16) $$\kappa ={\displaystyle \frac{P_o-P_e}{1-P_e}}$$where, $P_o$ = observed agreement and $P_e$ = expected agreement by chance

So the confusion matrix for test set is described in Table 3.

Table 3:

Confusion matrix for testset samples.

	Proposed model: Lesion (Square)	Proposed model: No lesion
Radiologist: Lesion (Circle)	50 (TP)	5 (FN)
Radiologist: No lesion	8 (FP)	237 (TN)

DOI: 10.7717/peerj-cs.2994/table-3

Observed agreement can be computed as

$$P_o={\displaystyle \frac{TP+TN}{Total}={\displaystyle \frac{50+237}{300}=\mathrm{0.9567.}}}$$

We can compute expected agreement $P_e$ as

P(Radiologist says Lesion) = (TP + FN)/Total = (50 + 5)/300 = 0.1833

P(Radiologist says No Lesion) = (FP + TN)/Total = (8 + 237)/300 = 0.8167

P(Model says Lesion) = (TP + FP)/Total = (50 + 8)/300 = 0.1933

P(Model says No Lesion) = (FN + TN)/Total = (5 + 237)/300 = 0.8067

$$\begin{array}{r}{P}_e=\left(P\left(\mathrm{L}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{b}\mathrm{y}\mathrm{R}\mathrm{a}\mathrm{d}\mathrm{i}\mathrm{o}\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{s}\mathrm{t}\right)\times P\left(\mathrm{L}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{b}\mathrm{y}\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{e}\mathrm{d}\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{l}\right)\right)\\ +\left(P\left(\mathrm{N}\mathrm{o}\mathrm{L}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{b}\mathrm{y}\mathrm{R}\mathrm{a}\mathrm{d}\mathrm{i}\mathrm{o}\mathrm{l}\mathrm{o}\mathrm{g}\mathrm{i}\mathrm{s}\mathrm{t}\right)\times P\left(\mathrm{N}\mathrm{o}\mathrm{L}\mathrm{e}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}\mathrm{b}\mathrm{y}\mathrm{P}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{s}\mathrm{e}\mathrm{d}\mathrm{M}\mathrm{o}\mathrm{d}\mathrm{e}\mathrm{l}\right)\right)\end{array}$$

$$P_e=\left(0.1833\times 0.1933\right)+\left(0.8167\times 0.8067\right)$$

$$P_e=0.0354+0.6588=\mathrm{0.6942.}$$

Therefore, Cohen’s Kappa $\kappa ={\displaystyle \frac{P_o-P_e}{1-P_e}={\displaystyle \frac{0.9567-0.6942}{1-0.6942}={\displaystyle \frac{0.2625}{0.3058}\approx \mathrm{0.8588.}}}}$

Hence, the Kappa values shows that there exists a strong agreement between radiologist and proposed model.

Matthews correlation coefficient

We can compute the Matthews correlation coefficient (MCC) (Matthews, 1975) using Eq. (17).

(17) $$MCC={\displaystyle \frac{TP\times TN-FP\times FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}.}$$

From Table 3, we can compute

$$MCC={\displaystyle \frac{50\times 237-8\times 5}{\sqrt{\left(50+8\right)\left(50+5\right)\left(237+8\right)\left(237+5\right)}}={\displaystyle \frac{11\text{,}850-40}{\sqrt{58\times 55\times 245\times 242}}}}$$

$$MCC={\displaystyle \frac{11\text{,}810}{\sqrt{58\times 55\times 245\times 242}}\approx {\displaystyle \frac{11\text{,}810}{\sqrt{190\text{,}237\text{,}700}}\approx {\displaystyle \frac{11\text{,}810}{13\text{,}793.4}\approx \mathrm{0.856.}}}}$$

So, MCC ≈ 0.856, indicating strong correlation.

Negative predictive value

Negative predicitve value (NPV) (Altman & Bland, 1994) is defined as

$$NPV={\displaystyle \frac{TN}{TN+FN}={\displaystyle \frac{237}{237+5}={\displaystyle \frac{237}{242}\approx 0.979}}}$$

So, NPV ≈ 97.9%, reflecting the model’s ability to correctly identify healthy cases.

Performance of proposed model using K-fold cross-validations

We have also performed experiments for K-fold cross-validations on Ayub Teaching Hospital dataset. We performed 5-fold cross-validations for our experimental study. In Fig. 11, the proposed model consistently achieves high accuracy, ranging from 92% to 96% across the folds, with an average of approximately 94% indicating model stability.

Performance comparison using baselines

Table 4 presents the comparison of the proposed model with state-of-the-art approaches, namely CNN (Koh et al., 2022), CNN-SL (Guan & Loew, 2017), and CNN-HD (Jalalian et al., 2017), based on accuracy and MAP. The baseline CNN achieved a 94.44% accuracy as MAP was 93.66%, while CNN-SL recorded 93.76% accuracy and a MAP of 94.56%. The CNN-HD improved to 95.01% accuracy and 94.33% MAP. Our CNBiFPN ends up exceeding all baselines, where we obtained 96.11% accuracy and a MAP of 97.03%. Furthermore, the results suggest that early detection of breast cancer using CT scan images with the CNBiFPN model was proven to be highly effective.

Table 4:

Comparison of results of proposed model with baselines.

Approach used	Accuracy	MAP
CNN (Koh et al., 2022)	94.44%	93.66%
CNN-SL (Guan & Loew, 2017)	93.76%	94.56%
CNN-HD (Jalalian et al., 2017)	95.01%	94.33%
CNBiFPN	96.11%	97.03%

DOI: 10.7717/peerj-cs.2994/table-4

Conclusion

Breast cancer continues to pose a significant global health challenge, and early detection remains a cornerstone for improving patient outcomes. Traditional imaging techniques like mammography, ultrasound, and MRI are widely utilized; however, they often face limitations such as high false-positive rates, radiation exposure, and difficulty in detecting tumors in dense breast tissues. CT imaging offers a promising alternative because it provides high-resolution, three-dimensional visualization of breast tissues, enabling more accurate identification of abnormalities. This article proposes a novel deep learning framework called the convolutional neural bidirectional feature pyramid network for the early detection and precise localization of breast cancer lesions in CT scan images. Unlike conventional convolutional neural network models, the CNBiFPN incorporates multi-scale feature extraction and bidirectional feature fusion, effectively capturing both fine-grained and high-level contextual information. This dual capability is critical for detecting tumors with varying shapes, sizes, and textures, which are common challenges in breast cancer imaging. The proposed framework underwent rigorous validation using a real-world dataset from the radiology department of Ayub Teaching Hospital Abbottabad, Pakistan. The experimental results demonstrated a classification accuracy of 96.11%, which outperforms many current state-of-the-art models and exhibits a 1.71% improvement over traditional baselines. An important aspect of the study was cross-verification of the model’s outputs with expert radiologists’ opinions, emphasizing that the model’s predictions align well with clinical expectations.