Osteosarcoma is the most common primary bone malignancy, arising most frequently in the rapidly growing metaphyseal regions of long bones in children and young adults. The distal femur, proximal tibia, and proximal humerus account for the majority of cases, though virtually any bone can be affected. Roughly 1,000 people are diagnosed each year in the United States, with the disease predominantly striking those between ages 10 and 30. Tumors account for approximately 2% of cancers in children under 14 and 3% of tumors in teenagers over 14.
Survival statistics: The five-year survival rate for children aged 0-14 diagnosed with osteosarcoma is 68%, and teenagers aged 15-19 share the same 68% rate. For patients across all ages whose cancer is detected before it spreads beyond its origin site, the five-year survival rate rises to 74%. When the cancer has metastasized to surrounding tissue, organs, or lymph nodes, survival at five years falls to 27%, underscoring the importance of early, accurate detection.
Diagnostic challenge: Conventional osteosarcoma detection requires expert pathologist review of hematoxylin and eosin (H&E)-stained tissue samples under the microscope, a process that is both time-consuming and highly dependent on individual expertise. CT and MRI imaging, while useful for clinical staging, cannot provide the cellular-level information needed to distinguish viable tumor cells from necrotic tissue and non-tumor regions. Whole slide images (WSIs) of histopathology specimens fill this gap by capturing high-resolution digital representations of glass slides, enabling computational analysis at the cellular level.
This paper from Sensors (2022) presents a deep learning pipeline that automates osteosarcoma detection on WSIs, achieving up to 99.3% classification accuracy while also addressing two persistent gaps in prior AI medical imaging systems: data privacy and distributed processing. The model integrates blockchain security, edge computing, and fog computing within an Internet of Medical Things (IoMT) framework.
Research on automated osteosarcoma detection has progressed through several generations of techniques. Early computer-aided detection (CAD) systems relied heavily on hand-crafted (HC) features, where domain experts manually defined image descriptors such as texture, shape, and intensity statistics, then fed these into classical classifiers. While straightforward to implement on modest datasets, HC feature approaches become brittle on complex histopathology images where cell morphology, tissue architecture, and staining variation interact in ways that are difficult to encode manually.
MRI and CT limitations: Early automated systems for osteosarcoma used CT and MRI images as their input. Nasor and Obaid applied a combination of K-means clustering, Chan-Vese segmentation, iterative Gaussian filtering, and Canny edge detection to segment osteosarcomas in MRI. Altameem used fuzzy ranking-based segmentation in X-rays followed by a deep neural network with the Levenberg-Marquardt learning algorithm. These approaches are limited by the fundamental inability of MRI and CT to capture cellular-level detail, including nuclear density and internal tumor structure, that are critical for distinguishing viable tumor tissue from necrotic regions.
CNN-based WSI classifiers: Mishra et al. (2018) developed a CNN architecture trained on osteosarcoma WSIs to classify images into three categories: viable tumors, non-viable tumors, and non-tumors. Their model, which was compared against AlexNet, VGGNet, and LeNet baselines, achieved 92.4% total accuracy after reducing patch size to 128 x 128 pixels due to memory constraints. Arunachalam et al. (2019) combined ML and deep learning, using K-means clustering and Otsu's multi-level thresholding for region-of-interest extraction, followed by a support vector machine (SVM) achieving 89.9% accuracy, and CNN-based classifiers achieving 93.3% (patch-level) and 91.2% (tile-level).
Transfer learning results: Anisuzzaman et al. (2021) applied VGG19 and InceptionV3, both pre-trained and adapted via transfer learning, to classify full WSI inputs at 375 x 375 pixels. VGG19 achieved 93.91% multiclass accuracy while InceptionV3 reached 78.26%. A 2022 study by Pan et al. introduced a Noise-Reducing Cross-Attention Fully Connected Fusion Learning (NRCA-FCFL) transformer, achieving 99.17% accuracy. None of these prior systems incorporated data privacy protections, IoMT connectivity, or distributed computing architectures for deployment at scale.
The dataset used in this study was assembled by clinical investigators at the University of Texas Southwestern Medical Center in Dallas, drawing from archival H&E-stained osteosarcoma specimens collected at the Children's Medical Center in Dallas between 1995 and 2015. It represents 50 patients treated over a two-decade span, providing longitudinal clinical diversity in terms of patient age, treatment response, and disease stage at the time of resection.
Class distribution: The original dataset contains 1,144 images at a resolution of 1,024 x 1,024 pixels each. These images are annotated into three pathological classes by two independent medical specialists: non-tumor (536 images, 47%), necrotic tumor (263 images, 23%), and viable tumor (345 images, 30%). The three-class structure is clinically significant because the proportion of viable versus necrotic tumor tissue after chemotherapy is the primary histological measure of treatment response, directly guiding surgical and adjuvant therapy decisions.
Data augmentation: The original 1,144-image dataset was insufficient for robust deep learning training, particularly given its class imbalance. The authors applied data augmentation techniques, balancing each class to 1,100 images per category, for a total augmented training pool across three classes. Augmentation approaches included rotation, flipping, and other spatial transformations that preserve pathological features while increasing dataset size and variance. For AlexNet input compatibility, images were resized to 227 x 227 pixels through the pre-processing stage.
The annotation process, involving two medical specialists reviewing each image for the dominant cancer type, provides a relatively high-quality ground truth, though the paper does not report inter-annotator agreement statistics or the formal annotation protocol used to resolve disagreements between the two reviewers.
The proposed system is organized as a five-layer architecture that addresses both the machine learning classification task and the real-world deployment challenges of data privacy, distributed computation, and clinical integration. The layers are: data layer, pre-processing layer, edge computing layer, fog computing layer, and testing layer. This layered design is motivated by the limitations of centralizing all computation on a single server, which creates bottlenecks in latency, introduces single points of failure, and concentrates sensitive patient data in a way that raises privacy and regulatory concerns.
Data and blockchain layers: The data layer collects histopathology images using Internet of Medical Things (IoMT) devices, which act as the interface between physical medical imaging hardware and the digital processing pipeline. Patient data, both raw and pre-processed, are stored in blockchain-secured private cloud environments. Separate blockchain-secured clouds are maintained for training data (clouds H, M, and N for each optimizer) and for test data. The blockchain mechanism ensures that stored records cannot be tampered with after writing, providing both data integrity verification and an auditable chain of custody for patient information.
Edge computing layer: The edge computing layer trains the AlexNet transfer learning model directly at or near the data source, avoiding the latency and bandwidth costs of transmitting large WSI files to a centralized server. Three separate instances of the model are trained in parallel, one with each of the three optimizers: stochastic gradient descent with momentum (SGDM), adaptive moment estimation (ADAM), and root mean squared propagation (RMSProp). Each trained model is then stored in its respective blockchain-secured private cloud before passing to the fog layer for selection.
Fog computing layer: The fog computing layer sits between the edge devices and the central cloud, providing additional computational capacity to compensate for the limited processing power of edge nodes. Based on training accuracy, the best-performing model among the three optimizer variants is selected at the fog layer and stored in a shared public model cloud for deployment. This selection process ensures that only the highest-quality model is used for patient-facing diagnosis without requiring edge devices to handle model comparison logic themselves.
The core classification engine is AlexNet, a convolutional neural network originally designed for large-scale image recognition tasks on the ImageNet dataset. Transfer learning is applied by reusing the convolutional layers that AlexNet learned from millions of natural images and fine-tuning the network on the osteosarcoma WSI dataset. This approach significantly reduces training time compared to training a CNN from random weight initialization, and it allows the model to leverage low-level visual features (edges, textures, gradients) learned from diverse image data even though the domain differs substantially from natural photographs.
Training configuration: All three optimizer experiments used identical hyperparameters: 50 epochs, 1,250 total iterations, 25 iterations per epoch, a learning rate of 0.001, and a single CPU for computation. The 70/30 train-test split resulted in approximately 2,310 training samples and 990 testing samples across the three balanced classes after augmentation. The MacBook Pro 2017 with 16 GB RAM and 512 GB SSD with integrated GPU was the hardware platform, indicating the model was designed to run on accessible consumer hardware rather than requiring specialized GPU clusters.
Convergence behavior: The three optimizers showed notably different convergence trajectories. SGDM reached convergence by the tenth epoch and remained stable through epoch 50. ADAM took longer, converging at epoch 30 but then maintaining stability. RMSProp never achieved stable convergence, exhibiting random training fluctuations throughout all 50 epochs. This instability in RMSProp training makes its results unreliable for clinical deployment despite achieving the same numerical accuracy as SGDM on the test set, because a model that does not converge cleanly may behave unpredictably on unseen inputs.
At reduced epoch counts, performance degraded substantially. At 30 epochs, SGDM reached only 98.88% PA. At 20 epochs, accuracy fell to 92.8% for SGDM, 91.5% for ADAM, and 90.06% for RMSProp, confirming that the full 50-epoch training schedule is necessary to achieve peak performance on this dataset size.
The final testing set consisted of 990 images (330 per class) drawn from the 30% held-out portion of the augmented dataset. Performance was evaluated across twelve statistical parameters: prediction accuracy (PA), classification miss rate (CMR), sensitivity (Sen), specificity (Spec), F1 score, positive predictive value (PPV), negative predictive value (NPV), false positive rate (FPR), false negative rate (FNR), likelihood positive ratio (LPR), likelihood negative ratio (LNR), and Fowlkes-Mallows index (FMI). This comprehensive reporting approach allows for nuanced evaluation beyond simple accuracy.
SGDM results: The SGDM optimizer achieved the best overall performance. Its confusion matrix showed 328 correct non-tumor predictions (3 misclassified), 330 correct viable tumor predictions (none misclassified in that class, with 5 misclassifications from viable), and 322 correct viable predictions (2 non-tumor misclassified as viable). Aggregated statistical parameters: PA 99.30%, CMR 0.70%, Sen 98.80%, Spec 99.55%, F1 98.95%, PPV 99.10%, NPV 99.40%, FPR 0.45%, FNR 1.20%, LPR 218.02, LNR 0.01%, FMI 98.95%.
ADAM results: ADAM performed slightly below SGDM. Its confusion matrix recorded 330 correct non-tumor, 330 correct viable tumor, and 323 correct viable predictions, with 5 non-tumor and 4 viable tumor misclassifications. Statistical parameters: PA 99.09%, CMR 0.91%, Sen 98.20%, Spec 99.55%, F1 98.65%, PPV 99.09%, NPV 99.10%, FPR 0.45%, FNR 1.80%, LPR 216.05, LNR 0.02%, FMI 98.65%.
RMSProp results: RMSProp achieved the same PA as SGDM at 99.30% and identical statistical parameters across all metrics. However, because its training showed persistent instability, the authors flag RMSProp as unsuitable for clinical deployment despite its test-set numbers matching SGDM. The likelihood positive ratio of 217-218 across both SGDM and RMSProp indicates that a positive prediction from the model is more than 200 times more likely to reflect a true positive than a chance finding, a strong indicator of discriminative power.
The authors compare their proposed model against six published osteosarcoma classification systems spanning 2017-2022. The comparison table evaluates each system on accuracy and on three infrastructure features absent from all prior systems: blockchain security, IoMT integration, and fog/edge computing. Accuracy alone ranges from 84% (Mishra et al. 2017, CNN) to 99.17% (Pan et al. 2022, NRCA-FCFL transformer), with the proposed model's 99.30% exceeding all prior published baselines by a small but meaningful margin.
Head-to-head comparisons: The 2017 Mishra et al. CNN achieved 84% accuracy on the same three-class task. The 2017 Arunachalam et al. system using K-means and a flood-fill algorithm reached 95.5%. The 2018 Mishra et al. model using AlexNet, LeNet, and VGGNet achieved 92%. Anisuzzaman et al.'s VGG19 (2021) reached 96%. Arunachalam et al.'s complex trees and SVM ensemble (2019) achieved 89.9%. Pan et al.'s NRCA-FCFL transformer (2022) reached 99.17%. The proposed AlexNet with SGDM achieves 99.30%, outperforming all prior benchmarks including the transformer-based approach.
Infrastructure differentiation: The uniquely notable aspect of this comparison is that no prior system implemented any of the three infrastructure features. Every existing method used no blockchain, no IoMT integration, and no fog or edge computing. The proposed model is the only system in this comparison that combines state-of-the-art classification accuracy with a privacy-preserving, distributed deployment architecture, making it the first to address both diagnostic performance and real-world deployment requirements simultaneously.
It is worth noting that the NRCA-FCFL transformer (99.17%) used a different methodology and may not have been evaluated on exactly the same dataset split, so direct numeric comparisons should be interpreted cautiously. The improvements over most baselines are nonetheless substantial, particularly given that SGDM achieved near-perfect specificity at 99.55%, which is critical in a cancer detection context where false positives drive unnecessary interventions.
Dataset size and generalizability: The study relies on a dataset of 1,144 original images from 50 patients at a single institution, with augmentation used to compensate for limited training data. While augmentation effectively increases apparent dataset size for training, it does not substitute for genuine biological and institutional diversity. All images originate from one center's archival samples spanning a specific 20-year period, meaning the model has not been tested on data from different staining protocols, different slide scanner hardware, or patient populations from geographically distinct centers. External validation on independent cohorts is needed before the system could be deployed in a new clinical environment.
AlexNet architecture constraints: AlexNet, designed in 2012, requires input images resized to 227 x 227 pixels. Downsampling from the original 1,024 x 1,024 pixel WSI tiles discards fine-grained cellular detail that may be diagnostically relevant, particularly for borderline viable/necrotic tumor cases. More recent CNN architectures (ResNet, EfficientNet, Vision Transformers) can process larger input resolutions and have substantially deeper feature hierarchies that may better capture the spatial complexity of osteosarcoma histopathology. The authors acknowledge an intent to investigate more computationally and performance-superior deep learning models in future work.
Annotation transparency: The dataset was annotated by two medical specialists, but the paper does not report inter-annotator agreement statistics (such as Cohen's kappa), the protocol used to resolve disagreements, or whether a third expert was used as a tiebreaker. Given that the classification of viable versus necrotic tumor regions is a known source of pathologist variability, the quality of the ground-truth labels is a potential limitation that affects model ceiling performance and reported accuracy metrics.
Future directions: The authors indicate plans to explore newer, computationally efficient deep learning models that could handle higher-resolution inputs and improve accuracy further, particularly for the viable/non-viable distinction that carries direct treatment implications. Integration with clinical outcome data, such as chemotherapy response rates and long-term patient survival, would allow future models to predict not just tissue classification but prognostic endpoints. Prospective clinical validation and federated learning across multiple pediatric oncology centers would strengthen generalizability and address the single-center dataset limitation.