Table 1 Summary of cancer classification studies (categorized by dataset).
From: Cross-platform multi-cancer histopathology classification using local-window vision transformers
Dataset | Author(s) | Classification methods | Accuracy | Limitations |
|---|---|---|---|---|
Wisconsin Breast Cancer Dataset (WBCD) | Hoque et al25. | XGBoost | 94.74% | Overfitting risks, limited to tabular features, and generalization issues |
Batool and Byun26 | Voting Ensemble (ETC + LightGBM + RC + LDA) | 97.6% | Limited data diversity; not tested on multimodal data | |
IDC + BreaKHis | Kaddes et al27. | Hybrid CNN-LSTM | 99.17%, 99.90% | High computational cost; lacks interpretability; generalizability untested |
Real-world Breast Cancer Dataset | Wani et al28. | CNN + LGBM + SHAP | 98.29% | High complexity; unclear dataset details |
DCE-MRI Sequences | Prinzi et al29. | Time-Series (Rocket, MultiRocket) | 85.2% | Moderate performance; time-series complexity |
MIAS, CBIS-DDSM, INbreast | Chakravarthy et al30. | Hybrid CNN (VGG16, VGG19, ResNet50, DenseNet121) | 98.7%, 97.7%, 98.8% | Class imbalance; low generalizability |
ISIC 2019 | Ozdemir and Pacal31 | Hybrid CNN (ConvNeXtV2 + Separable Attention) | 93.48% | Needs testing on other clinical datasets |
ISIC, HAM10000, PH2, BCN20000 | Hussain et al32. | ANN, SVM, RF, K-Means | – | Traditional methods, not multiclass |
Hermosilla et al33. | CNNs, SVMs, hybrid | – | Not nine-class capable | |
Hussein & Abdulazeez34 | SVM, KNN, CNN (review) | – | Bias in datasets: small training sets | |
Yang et al35. | Deep CNNs, GoogleNet, Transformers | Up to 97% | Lighting inconsistency; imbalance | |
Naeem et al36. | SNC-Net (InceptionV3) | – | Architecture not disclosed | |
Imran et al37. | EfficientNetB0 + ACO | – | Lacks novelty; not multiclass | |
LC25000 (Lung & Colon) | Hasan et al38. | Lightweight multi-scale CNN | – | Cannot classify the nine types |
Shahadat et al39. | 1D channel-based attention CNN | ~ 100% | No overfitting prevention | |
Mengash et al40. | CLAHE + MobileNet + DBN | – | Not nine-class capable | |
Uddin et al41. | CNN, DNN, VGG-16, ResNet-50 | 98.8% | Single dataset only | |
Khan et al42. | Hybrid CNN-RNN (LSTM) | 98.70% | Not compared with SOTA models | |
IQ-OTH/NCCD | Gulsoy et al43. | FocalNeXt (ConvNeXt + FocalNet) | 99.81% | Needs clinical validation |
Krishnamoorthy et al44. | AC-WGAN + HOA | 99.2% | High computational cost | |
Musthafa et al45. | CNN + SMOTE | 99.64% | Bias from synthetic oversampling | |
Colon Dataset (DS1) | Azar et al46. | CNNs + various optimizers | – | Low accuracy; not multiclass capable |
GDS3837 | Zheng et al47. | Gene profiling + XGBoost | – | Lower accuracy |
TCIA, TCGA | Sangeetha et al48. | Multimodal Fusion DNN (MFDNN) | – | Lower accuracy |
LIDC, LUNA16, JSRT | Gayap et al49. | CNNs, dual-path networks, ViTs | – | Survey only; no multiclass model |
Dawood et al50. | CenterNet (ResNet-34 + attention) | – | Limited data; no multiclass classification | |
General (47-study SLR) | Nakach et al51. | Multimodal Deep Learning Fusion | ~ 99% | Lacks standardization; interpretability issues |
Lung and Colon | Sobur et al52. | Dual-CNN | – | Architecture details missing |
Survey Lung Dataset | Wani et al53. | CNN + XGBoost | – | Binary only; not multiclass |
