Bone cancer is rare but deadly, accounting for a small fraction of all cancer diagnoses yet carrying a five-year survival rate of roughly 40% according to Cancer Research UK (2021). That low figure is partly explained by how bone tumors behave: benign forms remain asymptomatic for months or years, making early detection difficult, while malignant forms spread aggressively to other body regions if not caught at an early stage. Primary bone cancer originates in bone cells themselves, whereas secondary bone cancer represents metastasis from another site, and the distinction matters enormously for treatment planning.
Clinical detection pathways: Bone cancer is diagnosed through a combination of symptoms, histopathological examination, and imaging. Symptom-based detection is unreliable because many bone tumors produce no discomfort until they grow large enough to press on surrounding structures. Histopathology, while definitive, is invasive and often cannot distinguish malignancy until disease is well advanced. CT imaging occupies a middle ground: it is non-invasive, widely available, and can visualize abnormal bone density and structural changes before symptoms emerge, making it a natural target for automated analysis.
The role of CAD systems: Prior computer-aided detection (CAD) work had applied random forest classifiers to 79 CT images with 91.47% accuracy, and K-means segmentation to small three-image MRI and CT datasets, with surface-area estimates differing from radiologist measurements by only 0.34 to 1.75%. These results were promising but relied on hand-crafted features. Deep convolutional neural networks (CNNs) learn features automatically from raw image data and have demonstrated superior performance in many medical imaging tasks. This study asks a focused, clinically relevant question: which CNN architecture performs best for binary (normal vs. cancerous) classification of bone CT images, and how do they compare on both accuracy and computational efficiency?
The authors from Kongu Engineering College and Sri Eshwar College of Engineering (Tamil Nadu, India) benchmarked ten CNN architectures on a standardized dataset, providing a direct comparison rarely available in the literature. Their goal was to identify a model suitable for deployment in a clinical CAD workflow, balancing diagnostic accuracy against the processing time constraints of a busy radiology department.
The study assembled a dataset of 1,141 CT scan images drawn from two publicly available repositories: Radiopaedia (radiopaedia.org) contributed 730 scans, and the Cancer Imaging Archive (cancerimagingarchive.net) contributed 411 scans. Of the total, 530 images depict confirmed bone cancer cases and 511 images represent normal bone, yielding a near-balanced binary classification task. This balance is notable because many medical imaging datasets suffer from severe class imbalance, which can bias a model toward the majority class.
Cancer subtypes represented: The cancerous images cover three distinct bone tumor types: parosteal osteosarcoma (a low-grade malignancy arising on the outer surface of the bone, often seen in the distal femur), osteochondroma (the most common benign bone tumor, characterized by cartilage-capped bony projections), and enchondroma (a benign intramedullary cartilage tumor frequently found in the small bones of the hands). Including both benign and malignant subtypes increases the clinical realism of the dataset and tests whether the CNN can capture meaningful distinctions that go beyond simply detecting structural abnormality.
Dataset split: Images were divided 80% for training (approximately 913 images), 10% for validation (approximately 114 images), and 10% for testing (approximately 114 images). This conventional split provides a held-out test set that neither influences model training nor hyperparameter tuning, allowing for an unbiased estimate of generalization performance. The same split was applied identically across all ten CNN models to ensure fair comparison.
A key limitation acknowledged by the authors is the relatively modest dataset size. While 1,141 images is adequate for a comparative benchmarking study, large-scale deep learning models such as ResNet101 and VGG19 were originally designed for and pretrained on ImageNet, which contains millions of images across 1,000 classes. The authors used transfer learning to adapt these pretrained weights to the bone cancer task, reducing the data requirement, but the gap between pretraining distribution and CT imaging remains a relevant source of performance variability.
Before any CNN sees the images, a three-stage preprocessing pipeline is applied to isolate and highlight the cancer-affected region. This pipeline mirrors classical computer vision methodology and serves two purposes: it reduces the computational burden on the deep learning model by focusing attention on diagnostically relevant regions, and it removes noise artifacts that are inherent to CT acquisition.
Stage 1, Noise Reduction with Median Filtering: CT images contain salt-and-pepper noise introduced by photon statistics and detector electronics. The median filter, a non-linear method, replaces each pixel value with the median of its neighborhood rather than the average. Unlike linear filters such as Gaussian or Wiener filters, the median filter preserves sharp edges while eliminating impulse noise, making it well suited for CT images where boundary definition between normal and cancerous tissue is diagnostically important. The authors compared several filter types (average, median, Gaussian, Wiener) and found the median filter provided the best early-stage detection performance for bone cancer images specifically.
Stage 2, Tumor Segmentation with K-Means Clustering: K-means clustering is an unsupervised algorithm that partitions image pixels into k groups (clusters) based on intensity similarity. The algorithm initializes k centroids randomly, assigns each pixel to the nearest centroid by squared Euclidean distance, updates centroid positions by averaging assigned pixels, and iterates until convergence. For bone CT images, the cancer region typically exhibits higher radiodensity (appearing brighter on CT) than surrounding normal bone and soft tissue, so K-means can reliably isolate it as a distinct cluster. In the paper's figures, the cancer-affected region is highlighted in red on the segmented output, providing a clear visual boundary for subsequent edge detection.
Stage 3, Edge Detection with Canny Algorithm: Once the cancer region is segmented, the Canny edge detector extracts its precise boundaries. The Canny algorithm applies a Gaussian smoothing step, computes gradient magnitude using horizontal and vertical convolution kernels (detecting edges as intensity transitions), applies non-maximum suppression to thin detected edges to single-pixel width, and uses dual thresholding (a low and a high threshold derived from the image histogram) to distinguish strong from weak edges. Among Sobel, Prewitt, Roberts, and Canny methods tested, Canny produced the most reliable edge localization and noise resistance for bone cancer images. The result is a clean outline of the tumor boundary that can also serve as a region-of-interest mask for the downstream CNN.
The authors selected ten CNN architectures spanning multiple generations of deep learning design, all adapted for binary classification (normal vs. cancerous bone) with a final softmax output layer. All models shared identical training hyperparameters: input image size 227 x 227 pixels, batch size 32, Adam optimizer with learning rate 0.001, binary cross-entropy loss function, and 20 training epochs. This controlled comparison isolates the effect of architecture choice from training procedure.
AlexNet: The pioneering 2012 architecture that established CNNs as dominant in image classification. It uses eight layers: five convolutional layers (with filter sizes from 11x11 down to 3x3, starting with 96 filters and pooling to reduce spatial dimensions) followed by three fully connected layers ending in a 2-class softmax output. ReLU activation functions are applied after each convolutional layer. AlexNet is relatively shallow and computationally lightweight compared to later architectures, with a 29-minute training time in this study.
Residual Networks, ResNet50 and ResNet101: ResNets introduced skip connections (residual connections) that route the input of a block directly to its output, allowing gradients to flow through very deep networks without vanishing. ResNet50 has 50 layers and ResNet101 has 101 layers. Despite their theoretical depth advantage, both underperformed AlexNet in testing accuracy (81% and 89% respectively) while requiring substantially more compute time (50 and 71 minutes). The additional depth did not translate to better performance on this relatively small, single-domain dataset.
VGG Networks (VGG16 and VGG19): VGG architectures use very small 3x3 convolutional filters stacked in multiple consecutive layers, trading filter size for depth. VGG16 has 16 weight layers; VGG19 has 19. Both were the slowest models tested (120 and 150 minutes), with testing accuracies of only 74% and 80%. The high parameter count of VGG networks makes them computationally expensive and prone to overfitting on limited data.
EfficientNet B0 and B2: EfficientNet introduced compound scaling, simultaneously increasing network width, depth, and input resolution by a fixed ratio to maximize performance per parameter. EfficientNet B0 achieved 89% testing accuracy in the fastest time of all models (17 minutes), making it the most efficient architecture in the study. EfficientNet B2, a slightly scaled-up variant, achieved 91% accuracy in 48 minutes. DenseNet121, Xception, and InceptionV3 all underperformed on this dataset: DenseNet121 (68% testing accuracy, 33 minutes), Xception (68%, 105 minutes), and InceptionV3 (69%, 51 minutes). Dense connectivity and separable convolutions, which excel on large and diverse datasets, appeared to offer no advantage here.
The complete performance table tells a nuanced story about the relationship between architectural complexity, training time, and generalization. AlexNet achieved 98% training accuracy, 98% validation accuracy, and 100% testing accuracy, with a computational processing time of 29 minutes. Its training and validation loss curves converged at epoch 14, and the epoch count was fixed at 20 for all models based on the point when any model first reached 100% accuracy in testing. That model was AlexNet, which therefore set the benchmark epoch for the entire comparison.
Detailed results by architecture: ResNet50 produced training/validation/testing accuracies of 84/83/81%, requiring 50 minutes. ResNet101 achieved 88/92/89% in 71 minutes. VGG16 scored 83/77/74% in 120 minutes, and VGG19 scored 86/87/80% in 150 minutes. DenseNet121 reached only 64/64/68% in 33 minutes. EfficientNet B0 hit 86/94/89% in 17 minutes, and EfficientNet B2 scored 87/91/91% in 48 minutes. Xception achieved 65/58/68% in 105 minutes, and InceptionV3 scored 59/59/69% in 51 minutes.
Interpreting AlexNet's dominance: The superior performance of AlexNet on this specific dataset likely reflects several factors. First, AlexNet's shallower architecture (8 layers vs. 50-101 for ResNets) may be better matched to the relatively small dataset: deeper networks require more data to fill their parameter space without overfitting. Second, AlexNet's larger initial convolutional filter (11x11 at stride 4) can capture coarser spatial patterns that may be diagnostically relevant in CT bone images, such as the irregular, high-density masses characteristic of osteosarcoma. Third, the specific image preprocessing pipeline, which produced clean, segmented cancer regions, may have reduced the image complexity to a level where a simpler architecture suffices.
Accuracy vs. efficiency trade-off: EfficientNet B0's combination of 89% accuracy in only 17 minutes positions it as the strongest contender for real-time clinical deployment if perfect accuracy is not the sole criterion. AlexNet's 29 minutes and 100% accuracy remain the performance benchmark, but the gap between AlexNet and EfficientNet B0 on a larger, more diverse dataset may narrow considerably. VGG16 and Xception represent poor trade-offs: high computational cost with below-average accuracy, making them unsuitable for this application.
The counterintuitive finding that AlexNet outperformed deeper architectures like ResNet101 and VGG19 challenges the common assumption that more layers always yield better results. The explanation requires understanding the interplay between dataset size, architecture capacity, and the nature of bone CT image classification.
Overfitting and parameter-to-sample ratio: ResNet50 has approximately 25 million parameters; ResNet101 has around 44 million; VGG16 has 138 million. With only 913 training images, the parameter-to-sample ratio for VGG16 exceeds 150,000 parameters per image, a configuration virtually guaranteed to produce overfitting without aggressive regularization. AlexNet has roughly 60 million parameters, which may seem large, but its architecture concentrates these parameters in three fully connected layers (4096-4096-2 neurons) that are straightforward to regularize with dropout. The authors used dropout in their training configuration, which would particularly benefit the simpler fully connected structure of AlexNet.
Transfer learning effectiveness: All models were initialized with ImageNet pretrained weights. For natural image classification, ResNets and VGGNets have demonstrated clear superiority over AlexNet. But CT bone images differ fundamentally from natural photographs: they are grayscale (converted to 3-channel for input compatibility), high-contrast, and structured around radiodensity rather than color or texture. The hierarchical features learned by very deep networks from natural images may not transfer as cleanly to the specific edge and density patterns that discriminate bone cancer on CT. AlexNet's lower-level features, learned at 11x11 kernel scale, may better capture the coarse structural patterns in bone CT.
EfficientNet's efficiency advantage: EfficientNet B0's strong efficiency (89% accuracy, 17 minutes) derives from its use of depthwise separable convolutions and compound scaling, which reduce computation while maintaining representational capacity. Even though it did not match AlexNet's accuracy, EfficientNet B0's time advantage of 12 minutes per 20-epoch training run translates to substantially faster iteration cycles during model development and potentially faster inference in real-time clinical settings. For edge deployment on CT scanners with limited on-board compute, EfficientNet B0 may be the pragmatic choice.
The study's findings are compelling within its experimental scope, but several structural limitations prevent direct translation of these results into a deployed clinical tool. Understanding these constraints is essential for interpreting the 100% AlexNet test accuracy claim in context.
Dataset size and diversity: 1,141 CT images represents a small dataset by deep learning standards. The held-out test set contains approximately 114 images, meaning that a single misclassification represents roughly 0.9% test error and a single correct prediction represents 0.9% accuracy. AlexNet's 100% test accuracy therefore means it correctly classified all approximately 114 test images, which is impressive but represents a narrow statistical sample. The confidence interval around 100% accuracy on a ~114-image test set is wide, and performance on a 10,000-image validation set would be far more clinically meaningful.
Single-institution retrospective design: Both data sources are open repositories rather than prospectively collected institutional cohorts. The CT images were acquired across multiple scanners, institutions, and imaging protocols, which provides some diversity, but the dataset lacks the clinical metadata (patient age, cancer stage, tumor size, treatment history) needed to stratify model performance by clinically relevant subgroups. A model that works well across all comers may perform poorly on specific subpopulations, such as pediatric patients or those with secondary (metastatic) bone lesions rather than primary bone cancer.
Binary classification limitation: The models classify images as either normal or cancerous, with no further granularity. In clinical practice, the more pressing diagnostic challenges are distinguishing benign bone lesions (such as enchondroma or osteochondroma) from low-grade malignancies, and grading malignant tumors by histological type (osteosarcoma vs. chondrosarcoma vs. Ewing sarcoma). A binary normal/cancer classifier cannot assist with these distinctions, limiting its practical utility to a screening role rather than a diagnostic one.
Absence of external validation: The study splits a single dataset into training, validation, and test subsets. External validation, testing a model trained on one dataset against a completely independent dataset from a different institution or country, is the minimum standard for clinical AI tools. Without it, the reported 100% accuracy may reflect dataset-specific features (scanner type, patient demographics, acquisition protocol) rather than generalizable diagnostic capability. The authors acknowledge the need for larger, more diverse datasets in future work.
The study lays groundwork for an automated CT-based bone cancer detection pipeline, but the path from benchmark results to a clinically validated CAD system requires several additional steps that the authors identify, either explicitly or through the gaps in their current methodology.
Scaling the dataset: The most immediate need is a larger, more diverse CT dataset with confirmed histopathological diagnoses as ground truth. Federated learning approaches, where models train on distributed hospital datasets without centralizing patient data, offer a practical path to assembling the large multi-institutional datasets required for clinical validation while respecting patient privacy regulations. Multi-center datasets from sarcoma referral centers in different geographic regions would also address spectrum bias, ensuring the model encounters the full clinical diversity of bone cancer presentations.
Multi-class classification: Extending from binary (normal vs. cancer) to multi-class classification (distinguishing osteosarcoma, chondrosarcoma, Ewing sarcoma, osteochondroma, enchondroma, bone metastasis, and normal) would substantially increase clinical utility. This would require annotated datasets with histopathologically confirmed subtype labels and may benefit from specialized architectures such as attention-based networks that can focus on specific morphological features distinguishing subtypes, such as cortical breakthrough patterns in osteosarcoma versus the lobular calcified matrix of chondrosarcoma.
Multi-modal integration: CT alone captures bone structure and density well but provides limited information about tumor vascularity, soft tissue extension, and metabolic activity. Integration with MRI (which excels at soft tissue characterization and marrow involvement) and PET (which captures metabolic activity as a surrogate for malignant behavior) would create a richer input for classification. Multi-modal fusion architectures, either early fusion (combining raw images from different modalities) or late fusion (combining predictions from modality-specific networks), have shown promise in other cancer types and represent a natural extension of this work.
Explainability and clinical workflow integration: Deep learning models function as black boxes, which limits clinician trust and regulatory approval. Gradient-weighted class activation mapping (Grad-CAM) and similar explainability techniques can highlight which CT regions drove a classification decision, producing a visual overlay that radiologists can evaluate for anatomical plausibility. The authors suggest that future iterations of this pipeline should incorporate such explainability features, along with prospective validation in a simulated or real clinical environment where the model's output is compared against radiologist performance on the same cases.