Formalising the Architecture: ResNet and Predictive Morphology

As HistoHelper shifts from a conceptual MVP into a structured research tool, it is necessary to formally ground the architecture in established literature. The primary objective remains unchanged: to determine whether a Convolutional Neural Network (CNN) can identify morphological signatures within primary melanoma Whole Slide Images (WSIs) that predict specific organ tropism (e.g., pulmonary versus cerebral metastasis).

The Rationale for Residual Networks (ResNet)

Training deep CNNs from scratch on gigapixel histological data is computationally prohibitive and highly prone to vanishing gradients. To bypass this, HistoHelper utilises Transfer Learning via the ResNet18 architecture.

Introduced by He et al. (2016), Residual Networks resolved the degradation problem inherent in deep learning by implementing “skip connections.” Instead of requiring each successive layer to map an entirely new unreferenced function, skip connections allow the gradient to bypass layers, prompting the network to learn residual mappings (He et al., 2016). This architectural modification enables the training of exceptionally deep networks without the associated loss of accuracy or gradient disappearance (He et al., 2016).

For HistoHelper, a pre-trained ResNet18 model (pre-trained on the ImageNet database) provides a robust foundation for feature extraction. The terminal fully connected layer is replaced to accommodate a binary classification output (Primary vs. Metastatic), allowing the model to leverage its pre-existing capacity for spatial hierarchy detection while fine-tuning its weights specifically for cellular morphology.

Predictive Morphology in Oncology

The integration of AI in dermatopathology has demonstrated that CNNs can identify malignant melanoma from H&E-stained WSIs with accuracy comparable to expert pathologists (Dika et al., 2020). However, predicting where a primary lesion will metastasise represents a paradigm shift from diagnostic to prognostic AI.

Currently, our pipeline extracts 1120x1120 pixel patches from the SPIDER-Skin dataset. The DataLoader normalises these tensors using standard ImageNet parameters, applying morphological augmentations (rotations, flips, color jittering) to prevent overfitting during training.

By pushing these normalised tensors through the MPS-accelerated ResNet18, we aim to map abstract, sub-visual pixel relationships to known clinical outcomes. If the skip-connections within ResNet can reliably propagate features indicative of tropism, we move closer to personalised, proactive oncology.

References

• Dika, E., Campione, E., Lambertini, M., et al. (2020). Artificial intelligence in melanoma dermatopathology: A review of literature. Journal of Cutaneous Pathology, 47(11), 1047-1051. https://doi.org/10.1111/cup.13745
• He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 770-778).