Algorithm 1 for Lung Cancer Classification Model

Input: CT volume $V\in {\mathbb{R}}^{S\times H\times W}$, bounding boxes $\left\{B_s{\}}_{s=1}^S\right.$ Output: Predicted histological subtype $\widehat{y}$ 1. CT Volume Preprocessing 1.1 Intensity clipping: $\mathrm{For} \mathrm{each} \mathrm{slice} s=1\dots S,$ $\mathrm{perform} \mathrm{lung} \mathrm{window} \mathrm{intensity} \mathrm{clipping}\text{:} I_p^s(x,y)\leftarrow \mathrm{m}\mathrm{i}\mathrm{n}({\tau }_{\mathrm{m}\mathrm{a}\mathrm{x}},\mathrm{m}\mathrm{a}\mathrm{x}({\tau }_{\mathrm{m}\mathrm{i}\mathrm{n}},I^s(x,y)))$ 1.2 Normalization: $\mathrm{Normalize} \mathrm{clipped} \mathrm{slice} \mathrm{for} \mathrm{model} \mathrm{compatibility}\text{:} I_s^{\prime }(x,y)\leftarrow {\displaystyle \frac{I_p^s(x,y)-\mathrm{m}\mathrm{i}\mathrm{n}(I_p^s)}{\mathrm{m}\mathrm{a}\mathrm{x}(I_p^s)-\mathrm{m}\mathrm{i}\mathrm{n}(I_p^s)}}$ 1.3 RGB conversion:$\mathrm{Convert} \mathrm{normalized} \sin \mathrm{gle}\text{-}\mathrm{channel} \mathrm{slice} \mathrm{to} \mathrm{RGB}\text{:} I_s^{rgb}(x,y)\leftarrow [I_s^{\prime }(x,y),I_s^{\prime }(x,y),I_s^{\prime }(x,y)]$ Representation: $V\leftarrow \{I_s^{rgb}{\}}_{s=1}^S\in {\mathbb{R}}^{S\times H\times W\times 3}$ 2. Tumor Region Segmentation using MedSAM 2.1 Bounding box extraction: $\mathrm{Extract} \mathrm{bounding} \mathrm{box} \mathrm{for} \mathrm{slice} s$$\text{:} B_s\leftarrow (x_{\mathrm{m}\mathrm{i}\mathrm{n}},y_{\mathrm{m}\mathrm{i}\mathrm{n}},x_{\mathrm{m}\mathrm{a}\mathrm{x}},y_{\mathrm{m}\mathrm{a}\mathrm{x}})$ 2.2 MedSAM mask generation: $\mathrm{Forward} \mathrm{slice} \mathrm{and} \mathrm{bounding} \mathrm{box} \mathrm{to} \mathrm{MedSAM} \mathrm{to} \mathrm{generate} \mathrm{mask}\text{:} M_s\leftarrow f_{\mathrm{MedSAM}}(I_s^{rgb},B_s),M_s\in \{0,1{\}}^{H\times W}$ 2.3 Mask merging:$\mathrm{If} \mathrm{multiple} \mathrm{regions} \mathrm{exist}, \mathrm{merge} \mathrm{masks} \mathrm{pixel}\text{-}\mathrm{wise}\text{:} M_s(x,y)\leftarrow {\mathrm{m}\mathrm{a}\mathrm{x}}_k{M}_{s,k}(x,y)$ 2.4 Tumor slice selection:$\mathrm{Retain} \mathrm{tumor} \mathrm{slices}\text{:} \mathcal{T}\leftarrow \{s\mid {\sum }_{x,y}{M}_s(x,y)>0\}$ 3. Multi-Level Feature Extraction 3.1 Structural and Geometric Features 3.1.1 ORB keypoints: Detect ORB keypoints: $${\mathcal{K}}_s^{ORB}\leftarrow \{(x,y)\mid I_p-I_i>ϵ \mathrm{or} I_i-I_p>ϵ, i\in \{1,\dots ,16\left\}\right\}$$ Compute orientation using intensity centroid moments: $$\theta \leftarrow {\mathrm{t}\mathrm{a}\mathrm{n}}^{-1}\left({\displaystyle \frac{m_{01}}{m_{10}}}\right)$$ Generate rotation-invariant BRIEF descriptors: $$\tau (p;x,y)\leftarrow \left\{\begin{array}{cc}1& \mathrm{if} p(x)<p(y)\\ 0& \mathrm{otherwise}\end{array}\right.$$ 3.1.2 Harris corners: $\mathrm{Compute} \mathrm{Harris} \mathrm{corner} \mathrm{response}\text{:} R(x,y)\leftarrow \mathrm{d}\mathrm{e}\mathrm{t}(H)-k\cdot (\mathrm{trace}(H){)}^2,$ $\mathrm{extract} \mathrm{corners}\text{:} {\mathcal{K}}_s^H\leftarrow \{(x,y)\mid R(x,y)>\delta \}$ 3.1.3 Shape descriptors:$\mathrm{Compute} \mathrm{spatial} \mathrm{moments} \mathrm{of} \mathrm{mask} M_s$$\text{:} m_{pq}\leftarrow {\sum }_x{\sum }_y{x}^p{y}^q{M}_s(x,y),(\overline{x},\overline{y})=(m_{10}/m_{00},m_{01}/m_{00})$ Central and normalized moments:${\mu }_{pq}\leftarrow {\sum }_x{\sum }_y(x-\overline{x}{)}^p(y-\overline{y}{)}^q{M}_s(x,y),{\eta }_{pq}\leftarrow {\displaystyle \frac{{\mu }_{pq}}{{\mu }_{00}^{1+(p+q)/2}}}$ Hu invariants:$H_i^{\prime }\leftarrow -\mathrm{sign}(H_i)\mathrm{l}\mathrm{o}\mathrm{g}(\mid H_i\mid ),i=1..7$ Structural vector:$f_{\mathrm{structure}}^s\leftarrow [f_{\mathrm{Hu}}\parallel {\mathcal{K}}_s^{ORB}\parallel {\mathcal{K}}_s^H]$ 3.2 Texture Features 3.2.1 Quantization:$\mathrm{Quantize} \mathrm{intensities}\text{:} Q_s(x,y)\leftarrow \mathrm{quantize}(I_s^{\prime })$ 3.2.2 GLCM computation:$\mathrm{Compute} \mathrm{GLCM} P(i,j)$ $\mathrm{and} \mathrm{derive}\text{:} \mathrm{Contrast},\mathrm{Homogeneity},\mathrm{Energy},\mathrm{Correlation},\mathrm{Dissimilarity}$ 3.2.3 LBP computation:$\mathrm{Compute} \mathrm{LBP}\text{:} \mathrm{LBP}(p)\leftarrow {\sum }_{n=0}^{P-1}s(g_n-g_c)2^n,s(x)=1 \mathrm{if} x\ge \mathrm{0,0} \mathrm{otherwise}$ 3.2.4 Texture vector:$f_{\mathrm{texture}}^s\leftarrow [f_{\mathrm{GLCM}}\parallel f_{\mathrm{LBP}}]\in {\mathbb{R}}^{d_t}$ 3.3 Deep Semantic Features 3.3.1 ROI extraction:$\mathrm{Extract} \mathrm{tumor} \mathrm{ROI}\text{:} I_s^{roi}\leftarrow I_s^{rgb}\odot M_s$ 3.3.2 ResNet-50 feature extraction:$\mathrm{Forward} \mathrm{through} \mathrm{ResNet}\text{-}50, \mathrm{global} \mathrm{average} \mathrm{pooling}\text{:} f_{\mathrm{deep}}^s\in {\mathbb{R}}^{2048},f_{\mathrm{deep}}^s\leftarrow \mathrm{GAP}(F(I_s^{roi}))$ 4. Hybrid Feature Fusion 4.1 Handcrafted fusion:$f_{\mathrm{hand}}^s\leftarrow [f_{\mathrm{structure}}^s\parallel f_{\mathrm{texture}}^s]$ 4.2 Hybrid fusion:$f_{\mathrm{hybrid}}^s\leftarrow [f_{\mathrm{hand}}^s\parallel f_{\mathrm{deep}}^s](\mathrm{dimension} d)$ 5. Optimal Feature Selection using PSO 5.1 Particle initialization:$\mathrm{Initialize} \mathrm{particle} \mathrm{population} p_i\in \{0,1{\}}^d$ 5.2 Velocity and position update:$v_i\leftarrow w{v}_i+c_1{r}_1(pbes{t}_i-x_i)+c_2{r}_2(gbest-x_i),x_i\leftarrow x_i+v_i$ 5.3 Fitness evaluation and selection:$F\leftarrow \alpha \cdot \mathrm{Acc}-\beta {\displaystyle \frac{N_s}{N_f}},f_{\mathrm{opt}}^s\leftarrow \mathrm{select}(f_{\mathrm{hybrid}}^s)$ 6. Sequential Modeling and Classification 6.1 Feature sequence construction:$\mathcal{F}\leftarrow \{f_{\mathrm{opt}}^1,\dots ,f_{\mathrm{opt}}^T\}$ 6.2 LSTM forward:$h_t,C_t\leftarrow \mathrm{LSTMCell}(x_t,h_{t-1},C_{t-1})$ 6.3 Final representation:$h_T$ 6.4 Logits and SoftMax:$z\leftarrow W{h}_T+b,{\widehat{y}}_i\leftarrow {\displaystyle \frac{e^{z_i}}{{\sum }_j{e}^{z_j}}}$ Return:$\mathrm{Predicted} \mathrm{histological} \mathrm{subtype} \widehat{y}$