StochasticNet Radiomic Sequencers

• The paper introduces a radiomic sequencer that uses random graph-based convolutional layers to halve learnable parameters and enhance cancer detection performance.
• The method employs bipartite random graphs to construct sparse convolutional layers, improving regularization and reducing overfitting in imaging analysis.
• Empirical evaluation on the LIDC-IDRI dataset demonstrates superior sensitivity, specificity, and overall accuracy compared to traditional radiomic techniques.

StochasticNet radiomic sequencers are a class of convolutional neural architectures designed to discover custom radiomic features directly from medical imaging data, particularly for the purpose of cancer detection. In contrast to conventional radiomics methodologies that rely on pre-defined, hand-crafted feature sets, StochasticNet radiomic sequencers employ a graph-based convolutional paradigm where each layer's connectivity is governed by random graphs. This framework facilitates extraction of feature representations tailored for characterizing unique cancer tissue phenotypes, improving both parameter efficiency and generalization in radiomics-driven classification tasks (Shafiee et al., 2015).

1. Graph-Based Convolutional Layer Construction

The fundamental innovation in StochasticNet radiomic sequencers is the representation of each convolutional layer as a bipartite random graph, where connections between input channels and output feature maps are stochastically instantiated. For each convolutional layer $l$, with $C_{l-1}$ input channels and $F_l$ output filters (each having spatial support $s \times s$), an adjacency-mask tensor $$A^l \in \{0,1\}^{C_{l-1} \times s \times s \times F_l}$$ is generated with entries $A^l_{c,u,v,k} \sim \mathrm{Bernoulli}(p)$, where $p = 0.5$. This results in each possible edge (connection) between input and output nodes being present with fixed probability, thereby inducing sparsity in the weight tensors $W^l$. Only those weights with corresponding $A^l_{c,u,v,k}=1$ are instantiated and learned; other positions remain permanently zeroed. This stochastic parameter reduction cuts the number of learned weights per convolutional filter in half, in expectation, compared to fully dense layers.

2. Forward Propagation and Feature Extraction

Within a stochastic convolutional layer, the computation for output feature maps $X^l \in \mathbb{R}^{F_l \times H_l \times W_l}$ given input $$X^{l-1} \in \mathbb{R}^{C_{l-1} \times H_{l-1} \times W_{l-1}}$$ is:

$$X^l_{k}[i,j] = \phi\left( \sum_{c=1}^{C_{l-1}} \sum_{u=1}^s \sum_{v=1}^s A^l_{c,u,v,k} W^l_{c,u,v,k} X^{l-1}_{c}\left[i+u-\left\lfloor \frac{s}{2} \right\rfloor,\, j+v-\left\lfloor \frac{s}{2} \right\rfloor\right] + b^l_k \right)$$

where $\phi(\cdot)$ denotes the layer’s nonlinearity (the specific function is not stated in the source), and $b^l_k$ denotes the bias term for filter $k$. This random masking acts as a form of structural regularization, potentially reducing overfitting and computational burden without diminishing representational power.

3. Radiomic Sequencer Architecture and Parameterization

The full StochasticNet radiomic sequencer is composed of three consecutive stochastic convolutional layers:

Layer	Filters ($F_l$)	Filter Size	Input Channels ($C_{l-1}$)	Connectivity ($p$)
Layer 1	32	$5 \times 5$	1	0.5
Layer 2	32	$5 \times 5$	32	0.5
Layer 3	64	$5 \times 5$	32	0.5

The output feature maps of the final convolutional layer are flattened into a $d$-dimensional radiomic feature vector $\mathbf{x} \in \mathbb{R}^d$, where $d = F_3 \times H_3 \times W_3$. The overall sequencer is represented by the composite mapping $\mathbf{x} = f_\theta(I)$, where $I$ is the input CT image patch and $\theta$ encompasses all layer-specific adjacency masks, weights, and biases.

4. Training Objective and Regularization Effects

Supervised training is performed for binary classification, predicting lesion malignancy (benign vs. malignant). The radiomic feature vector $\mathbf{x}$ is passed through a fully connected output layer and a two-class softmax for probability estimation. The model is trained to minimize the cross-entropy loss:

$$\mathcal{L}(\theta) = -\frac{1}{N} \sum_{n=1}^N \sum_{c=1}^2 \mathbf{1}\{y^{(n)} = c\} \log \hat{y}_c^{(n)}$$

where $(I^{(n)}, y^{(n)})$ denote training samples and corresponding binary labels, and $\hat{y}_c^{(n)}$ are the predicted class probabilities. The random sparsification of weights induced by the adjacency masks itself serves as implicit regularization, functioning analogously to DropConnect by preventing parameter co-adaptation, reducing overfitting, and potentially accelerating convergence. Optimizer selection, learning rate, and explicit regularization hyperparameters are not specified in the source.

5. Inference and Post-Processing

At inference time, each candidate lesion patch $I$ is processed by the fixed-structure radiomic sequencer, yielding the feature vector $\mathbf{x} = f_\theta(I)$. Probability estimates for the two classes are obtained via the softmax output; decision-making proceeds by thresholding or taking the argmax of these probabilities. No post-processing procedures beyond this softmax-based scoring are reported.

6. Empirical Evaluation on LIDC-IDRI

The StochasticNet radiomic sequencer was evaluated on the LIDC-IDRI dataset, comprising 42,340 lung lesion patches (including rotation-augmented samples) from 93 patients. Its classification performance was benchmarked against:

Method	Sensitivity (%)	Specificity (%)	Accuracy (%)
Belief Decision Trees (BDT) [Zinovev ’11]	–	–	54.32
Deep Autoencoding Radiomic Sequencer (DARS) [Kumar ’15]	83.35	20.18	75.01
StochasticNet Radiomic Sequencer (SNRS)	91.07	75.98	84.49

Area under the ROC curve (AUC) is not reported. The SNRS approach demonstrates a marked improvement over both traditional machine learning and deep autoencoding radiomics baselines, achieving high sensitivity, specificity, and overall accuracy on the augmented lung CT patch diagnosis task (Shafiee et al., 2015).

7. Theoretical and Practical Implications of Stochastic Connectivity

The stochastic connectivity paradigm confers several advantages:

• Parameter Efficiency: Each filter’s parameters are halved, reducing memory and computational requirements.
• Regularization: Structural sparsity introduced by random graph connectivity reduces the risk of overfitting by limiting direct co-adaptation between weights.
• Theoretical Validity: Random graph theory indicates that provided the connectivity probability $p$ is not too small, the resulting sparse network approximates the representational capabilities of a dense network.
• Empirical Robustness: The architecture yields strong classification results on large-scale, real-world clinical imaging datasets without the need for hand-crafted features or complex post-processing.

A plausible implication is that random-graph based convolutional architectures could be advantageous in other domains where high-dimensional medical data and modest sample sizes make overfitting a critical concern.

---

For foundational exposition and empirical benchmarks of the StochasticNet radiomic sequencer, see (Shafiee et al., 2015).