IQOTHNCCD Lung Cancer Dataset

• IQOTHNCCD Lung Cancer Dataset is a publicly available, expert-annotated CT image repository that supports deep learning research in lung cancer classification and explainability.
• The dataset comprises approximately 1,100 thoracic CT-slice images labeled as Normal, Benign, or Malignant, enabling robust evaluation of classification models.
• Advanced preprocessing, augmentation, and explainable AI techniques such as Grad-CAM and SHAP are applied to enhance model performance and clinical interpretability.

The IQOTHNCCD Lung Cancer Dataset is a publicly available, expert-annotated thoracic computed tomography (CT) image dataset developed to support research in automated lung cancer diagnosis, classification, and explainable AI. It forms the basis for a significant volume of contemporary research on deep learning–enabled lung cancer detection, providing multi-class ground truth (Normal, Benign, Malignant) and substantial methodological transparency. The dataset originates from the International Qatar–Oman Thoracic Hospital (IQ-OTH) and National Cancer Care & Research Centre (NCCD), and is widely disseminated in both raw (DICOM/PNG) and preprocessed forms (Abdollahi, 2023, Islam et al., 3 Dec 2025, Rai et al., 13 Aug 2025).

1. Dataset Composition and Labeling

The IQOTHNCCD dataset comprises 1,097–1,197 thoracic CT-slice images collected in 2019, with class partitioning into Normal, Benign, and Malignant as confirmed by board-certified radiologists and oncologists. The precise patient count is not universally reported; one study specifies acquisition from 110 patients (40 malignant, 15 benign nodules, 55 normal), while others focus on image-level summary statistics.

The following table classifies dataset structure as reported by major studies:

Study [arXiv ID]	Total Images	Normal	Benign	Malignant	Patients
(Abdollahi, 2023)	1,190	55	15	40	110
(Rai et al., 13 Aug 2025, Islam et al., 3 Dec 2025)	1,097/1,197	416	120	561	–

Image annotation involved clinical review with class assignment based on radiological and oncological consensus. Labels were assigned at the CT slice (2D) level in a three-class (Normal, Benign, Malignant) schema. There is no report of interobserver agreement statistics or multi-reader adjudication in the public documentation (Rai et al., 13 Aug 2025).

2. Image Acquisition, Preprocessing, and Augmentation

The CT acquisitions were obtained on a Siemens SOMATOM system with an imaging protocol specifying 1 mm slice thickness, breath-hold at full inspiration, and lung windows (width: 350–1,200 HU, center: 50–600 HU) (Abdollahi, 2023). Original data are stored as DICOM or PNG.

Preprocessing commonly consisted of:

• Down-sampling or resizing: Images were scaled to either 28×28 pixels (LeNet models (Abdollahi, 2023)) or 256×256 pixels (transfer learning studies (Islam et al., 3 Dec 2025, Rai et al., 13 Aug 2025)), using bilinear interpolation.
• Gray-to-RGB conversion: For compatibility with standard CNN backbones, single-channel (grayscale) images were channel-replicated to achieve 3-channel (RGB) input (Islam et al., 3 Dec 2025).
• Normalization: Either rescaling to 0,1 or mean/standard deviation normalization using ImageNet statistics was applied (Rai et al., 13 Aug 2025, Islam et al., 3 Dec 2025).
• Augmentation:
  • Random rotations (typically ±15–25°)
  • Random horizontal/vertical flips
  • Random shifts
  • Synthetic Minority Over-Sampling Technique (SMOTE) for class balancing in training partitions (Islam et al., 3 Dec 2025)
• No segmentation or explicit lung field localization was performed.

3. Data Partitioning and Experimental Protocols

Splitting protocols to assess generalization performance overwhelmingly employ a stratified partition:

• Train-Test Splits: Typical splits are ~80% for training, 20% for held-out test evaluation, with stratification to maintain class proportions (Islam et al., 3 Dec 2025). One study refers to a 85:15 train:validation ratio (Abdollahi, 2023).
• No prescribed k-fold or cross-validation: Experiments primarily rely on single stratified splits with early stopping on validation loss (Rai et al., 13 Aug 2025, Islam et al., 3 Dec 2025).
• Class Imbalance Handling: Addressed via weighted loss functions (inverse class frequency) or oversampling (SMOTE) (Rai et al., 13 Aug 2025, Islam et al., 3 Dec 2025).
• Label Distribution: Malignant class represents the majority, Benign is generally the minority. Test set supports: Normal ≈83, Benign ≈24, Malignant 113 (Islam et al., 3 Dec 2025).

4. Deep Learning Methods and Model Architectures

Research leveraging the IQOTHNCCD dataset benchmarks both classic and advanced CNN architectures:

• LeNet-5–style CNN: Adapted to three classes, using two convolutional layers (5×5 kernels), average pooling, and three fully connected layers with ~44,000 parameters. Both cross-entropy and focal loss are evaluated for optimal handling of class imbalance (Abdollahi, 2023).
• Custom CNN and Transfer Learning: Studies deploy custom architectures and fine-tune DenseNet121/169, ResNet152, VGG19 with new classifier heads. Frozen backbone weights are standard, with only the classifier trained (Rai et al., 13 Aug 2025).
• DenseNet169 + Squeeze-and-Excitation (SE) and Feature Pyramid Network (FPN): Channel attention (SE blocks) and multi-scale fusion (FPN) are integrated for robust feature extraction, with focal loss for imbalance mitigation. Output shape after last dense block: ~7×7×1664 (Islam et al., 3 Dec 2025).
• MobileNetV2 Feature Extractor + SVM: Deep features pooled and used with a linear SVM, employing StandardScaler normalization, yielding competitive results (Islam et al., 3 Dec 2025).

5. Quantitative Performance and Evaluation Metrics

Reported metrics align with clinical and machine learning standards: overall accuracy, class-specific precision/recall/F1, sensitivity, specificity, area under the ROC curve (AUC), and confusion matrices. Definitions conform to standard practice:

$$\text{Accuracy} = \frac{\text{TP} + \text{TN}}{\text{TP} + \text{TN} + \text{FP} + \text{FN}}$$

$$\text{Precision}_c = \frac{\text{TP}_c}{\text{TP}_c + \text{FP}_c} \qquad \text{Recall}_c = \frac{\text{TP}_c}{\text{TP}_c + \text{FN}_c}$$

$$F_{1,c} = 2\,\frac{\text{Precision}_c \times \text{Recall}_c}{\text{Precision}_c + \text{Recall}_c}$$

Key reported results (test set or validation):

Model/Study	Accuracy (%)	Sensitivity (%)	Specificity (%)	Precision (%)	F1 (%)
LeNet + focal loss (Abdollahi, 2023)	97.88	93.14	95.91	–	–
DenseNet169+SE+FPN (Islam et al., 3 Dec 2025)	98	–	–	Benign: 92	Benign: 94
ResNet152 (fine-tuned) (Rai et al., 13 Aug 2025)	97.3	–	–	86	91
DenseNet121 (fine-tuned) (Rai et al., 13 Aug 2025)	89.15	–	–	92	91
Custom CNN (Rai et al., 13 Aug 2025)	92.86	–	–	78	73
SVM+MobileNetV2 (Islam et al., 3 Dec 2025)	98	–	–	Benign: 95	Benign: 91

A “success percentage” of 99.51% for LeNet is cited in (Abdollahi, 2023) but not explained outside the abstract. Comparative results indicate the newer hybrid/explainable frameworks match or improve upon earlier approaches.

6. Explainable AI and Interpretability

Recent work emphasizes deployment of explainable AI frameworks using IQOTHNCCD:

• Grad-CAM: Class activation maps computed as $$L_{\mathrm{GradCAM}}^k = \mathrm{ReLU}\left(\sum_l \alpha_l^k A^l\right)$$, with $\alpha_l^k$ computed from class logit gradients. These maps overlay regions of a CT slice most influential for a given class assignment (Malignant, Benign, Normal), enhancing clinical transparency (Islam et al., 3 Dec 2025).
• SHAP (Shapley Additive Explanations): Applied both at the pixel (for neural models) and feature (for SVM) levels, SHAP decomposes the prediction into additive contributions of each input feature or pixel. The Shapley value formula is:

$$\phi_i = \sum_{S \subseteq \{1…M\} \setminus \{i\}} \frac{|S|!\,(M-|S|-1)!}{M!}\left[f_{S\cup\{i\}}(x_{S\cup\{i\}}) - f_S(x_S)\right]$$

This technique clarifies which anatomical structures in a CT slice drive Malignant vs. Benign or Normal predictions, supporting adoption in real-world medical practice (Islam et al., 3 Dec 2025, Rai et al., 13 Aug 2025).

7. Limitations and Research Use Cases

• Technical Limitations:
  • No CT scanner manufacturer or protocol metadata is universally available. Detailed acquisition parameters (beyond Siemens SOMATOM for 110-patient subset) are not reported in all studies (Rai et al., 13 Aug 2025).
  • Slice-level labels without lesion-level bounding boxes or segmentation masks.
  • Dataset size (≈1,100 images, ~110 patients) is modest by contemporary deep learning standards; Benign class is notably underrepresented.
  • No prescribed cross-validation; single splits may overstate generalizability.
• Suggested Applications:
  • Benchmarking multi-class (Normal/Benign/Malignant) thoracic CT slice classification.
  • Development of class-imbalance-resilient loss functions (e.g., focal loss).
  • Evaluation of explainable AI techniques in medical imaging (Grad-CAM, SHAP).
  • Comparative analysis of custom CNNs vs. state-of-the-art transfer learning models.
  • Radiomics and feature analysis for automated screening in resource-limited settings.

The clinical-grade annotation and widespread adoption of the IQOTHNCCD dataset have positioned it as a pivotal testbed for algorithmic advances in deep learning–driven lung cancer research, with continued evolution in both model performance and interpretability (Abdollahi, 2023, Rai et al., 13 Aug 2025, Islam et al., 3 Dec 2025).