Algorithm 2. Residual MLP Classifier

Input: Fused CNN embedding $x\in {\mathbb{R}}^{3328}$

Output: Softmax probability vector $\widehat{y} \in {\mathbb{R}}^2$

Input transformation

$h_0\leftarrow {\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}}_{0.3}\left[\mathrm{B}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\right(\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}\left({\mathrm{D}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{e}}_{1024}\right(x)+b_0)\left)\right)]$

2.Residual blocks

For each residual block k = 1,2:

$res\leftarrow h_{k-1}$

$h\leftarrow {\mathrm{D}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{e}}_{256}\left(h_{k-1}\right)\to \mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}\to \mathrm{B}\mathrm{a}\mathrm{t}\mathrm{c}\mathrm{h}\mathrm{N}\mathrm{o}\mathrm{r}\mathrm{m}\to {\mathrm{D}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{e}}_{1024}$

$h_k\leftarrow {\mathrm{D}\mathrm{r}\mathrm{o}\mathrm{p}\mathrm{o}\mathrm{u}\mathrm{t}}_{0.2}\left[\mathrm{R}\mathrm{e}\mathrm{L}\mathrm{U}\right(h+\mathrm{r}\mathrm{e}\mathrm{s}\left)\right]$

3.Output layer

$\widehat{y}\leftarrow \mathrm{S}\mathrm{o}\mathrm{f}\mathrm{t}\mathrm{m}\mathrm{a}\mathrm{x}\left({\mathrm{D}\mathrm{e}\mathrm{n}\mathrm{s}\mathrm{e}}_2\right(h_2\left)\right)$

4.Training configuration

The residual MLP was trained using Adam with cosine annealing, categorical cross-entropy with label smoothing of 0.1, a batch size of 32, and 25 epochs under stratified five-fold cross-validation.