IQ-OTH/NCCD CT Dataset for Lung Cancer Detection

• IQ-OTH/NCCD dataset is a publicly available resource of thoracic CT scans, curated to support lung cancer detection and explainable AI research.
• The dataset comprises 1,190 to 1,336 de-identified CT slices labeled as normal, benign, or malignant, with stratified splits ensuring balanced evaluation.
• Images undergo preprocessing including resizing, intensity normalization, and augmentation, while evaluation metrics such as accuracy, sensitivity, and AUC benchmark model performance.

The IQ-OTH/NCCD dataset is a publicly available thoracic computed tomography (CT) resource collected for the development and evaluation of computer-aided lung cancer detection and classification models. Assembled through a clinical collaboration between the Iraq Oncology Teaching Hospital (IQ-OTH) and the National Center for Cancer Diseases (NCCD), the dataset provides expertly annotated, de-identified CT slices across normal, benign, and malignant categories, enabling robust benchmarking of machine learning and explainable AI (XAI) approaches in oncologic imaging (Panboonyuen, 19 Jan 2026, Abdollahi, 2023, Rai et al., 13 Aug 2025, Chaddad et al., 2 Jul 2025).

1. Dataset Origin, Purpose, and Scope

The IQ-OTH/NCCD dataset was collected under routine clinical protocols at IQ-OTH and NCCD, primarily from January 2018 to December 2020 (with some studies focused on the Fall 2019 interval). Its explicit design goal is to facilitate both early detection/classification of lung cancer and the rigorous evaluation of interpretable AI methods in high-stakes diagnostic contexts (Panboonyuen, 19 Jan 2026). The dataset is intended to reflect real-world diagnostic challenges, incorporating three diagnostic classes: normal lung parenchyma, benign nodules, and malignant lesions.

All CT slices were reviewed and labeled by board-certified oncologists and radiologists. The clinical population comprises 110 unique patients drawn from multiple provinces and occupational backgrounds within Iraq, providing a degree of demographic and case diversity. The dataset release is fully de-identified, has institutional review board approval, and waives written consent via ethical oversight (Abdollahi, 2023).

2. Composition, Class Distributions, and Demographics

The core IQ-OTH/NCCD dataset consists of 110 subjects, each with a full CT study comprising 80–200 axial slices, for a total of approximately 1 190 unique DICOM images (Panboonyuen, 19 Jan 2026, Abdollahi, 2023). In augmented or expanded splits used for deep learning experiments, the sample may include up to 1 336 images (Chaddad et al., 2 Jul 2025), or 1 197 axial CT slices as reported elsewhere (Rai et al., 13 Aug 2025). The dataset strictly uses a per-slice, not volumetric (3D), classification paradigm.

Class and split distributions (as reported in various studies):

Split	Benign	Malignant	Normal	Total
Training	287	449	334	1,070
Validation	36	56	41	133
Test	36	56	41	133
Overall	359	561	416	1,336

Original, non-augmented figures: 40 malignant, 15 benign, 55 normal cases across 1 190 images (Abdollahi, 2023). Another account reports: Normal = 416, Benign = 120, Malignant = 561, total = 1 197 (Rai et al., 13 Aug 2025). This suggests that minor fluctuations in sample size across studies are due to pre-processing, augmentation, or exclusion of incomplete slices.

Demographic and clinical metadata includes anonymized patient identifiers, with varied age and gender distributions; however, explicit age/gender counts and advanced clinical features (e.g., staging, smoking history) are not provided (Panboonyuen, 19 Jan 2026, Chaddad et al., 2 Jul 2025).

3. Imaging Protocols, File Formats, and Preprocessing

Acquisition was performed on Siemens SOMATOM CT systems. Standardized parameters include:

• Tube voltage: 120 kV
• Slice thickness: 1 mm (in some cases, 1.0–1.25 mm typical; in-plane ~0.6 mm × 0.6 mm)
• Window width: 350–1 200 HU
• Window center: 50–600 HU
• Breath-hold: full inspiration
• Images are non–contrast-enhanced, single-channel (grayscale)

De-identification removed all protected health information from DICOM headers. Image preprocessing varies by publication:

• For deep neural network input: Slices are resized (often to 224×224 or 256×256 pixels), intensity normalized (either to zero mean/unit variance per slice or scaled via ImageNet statistics), and in some pipelines, center cropped (Panboonyuen, 19 Jan 2026, Rai et al., 13 Aug 2025, Chaddad et al., 2 Jul 2025).
• Augmentation: Pipelines include random rotations (±15°), horizontal flips ($p=0.5$), affine transforms, contrast changes, added noise, and Gaussian blur (Chaddad et al., 2 Jul 2025, Rai et al., 13 Aug 2025). Some studies employ minimal augmentation, relying on stratified splits (Panboonyuen, 19 Jan 2026). LeNet-based studies downsample input to 28×28 pixels (Abdollahi, 2023).

No explicit lung segmentation or 3D contextualization is applied; each slice is processed independently.

4. Annotation Protocols and Ground-Truth Definition

Expert labeling was conducted by experienced oncologists and radiologists. Each axial slice was categorized as either normal, benign, or malignant based on standardized radiologic criteria. For malignant lesions, pixel-level ground-truth masks ($M_i$) delineating tumor regions are available in certain versions, enabling quantitative localization evaluation (such as overlap with Grad-CAM saliency maps) (Panboonyuen, 19 Jan 2026).

Consensus review was performed among at least two experts; formal inter-rater reliability measures (e.g., ICC, κ statistics) are not reported. The label set is inherited from primary annotations, and no details are provided regarding voting or adjudication protocols in later studies (Panboonyuen, 19 Jan 2026, Rai et al., 13 Aug 2025).

5. Data Splits, Class Imbalance Management, and Preprocessing Pipelines

Stratified splits are standard: a 60/20/20 approach (train/val/test) is explicitly stated in some publications (Panboonyuen, 19 Jan 2026), while others use 80/10/10 with explicit image counts and fixed random seed for reproducibility (Chaddad et al., 2 Jul 2025). Some studies report only that "stratified splits" were used and leave proportions unspecified (Rai et al., 13 Aug 2025), though splits are always balanced across the three classes.

Class imbalance is nontrivial, with Malignant slices outnumbering Benign (typically by $\approx4.7\times$) and Normal ($\approx1.3\times$), motivating cost-sensitive loss weighting ($w_c \propto \frac{1}{N_c}$) and on-the-fly data augmentation for underrepresented labels (Rai et al., 13 Aug 2025, Chaddad et al., 2 Jul 2025).

Preprocessing steps include: intensity normalization ($I' = \frac{I - \mu}{\sigma}$ on a per-slice basis), resizing, conversion to the required tensor/channel format, and, in some cases, per-pixel mask alignment for localization tasks (Chaddad et al., 2 Jul 2025, Panboonyuen, 19 Jan 2026).

6. Evaluation Metrics and Benchmarking

Classification performance is quantitatively assessed via:

• Accuracy ($\frac{TP+TN}{TP+FP+FN+TN}$)
• Sensitivity ($\frac{TP}{TP+FN}$)
• Specificity ($\frac{TN}{TN+FP}$)
• Precision ($\frac{TP}{TP+FP}$)
• F1 score ($M_i$0)
• Area Under ROC Curve (AUC), generally "one-vs-all" for multiclass (Chaddad et al., 2 Jul 2025, Rai et al., 13 Aug 2025)

For segmentation/explainability-based studies: localization accuracy

$M_i$1

faithfulness,

$M_i$2

and explanation consistency

$M_i$3

are reported (Panboonyuen, 19 Jan 2026).

Published benchmarks include:

Model	Accuracy	Precision	Recall	F1	AUC
ResNet50+	99.25%	99.21%	99.07%	99.13%	0.999
DenseNet121	≤92%	92%	90%	91%	n/a
ResNet152	97.3%	n/a	n/a	n/a	n/a
LeNet-5+focal	97.88%	n/a	93.14%	n/a	n/a

Studies highlight the significant impact of architecture selection, augmentation policy, and class weighting on achieved metrics (Chaddad et al., 2 Jul 2025, Rai et al., 13 Aug 2025, Abdollahi, 2023).

7. Limitations, Usage Considerations, and Recommendations

Documented limitations include moderate overall sample size (1 197–1 336 slices), single-center or limited-institutional origin, restricted patient covariates, and lack of 3D volumetric context (Chaddad et al., 2 Jul 2025, Rai et al., 13 Aug 2025). Detailed acquisition parameters are sometimes incomplete; users are advised to consult the original Mendeley Data repository for further reporting. Annotation provenance, such as number of annotators per slice and inter-labeler agreement, is not systematically published (Rai et al., 13 Aug 2025).

The exclusive use of 2D slices precludes volumetric lesion assessment. The dataset is widely used for benchmarking supervised lung cancer classifier architectures, as well as for the evaluation of XAI pipelines (e.g., Grad-CAM, SHAP) in clinical imaging. However, caution is advised regarding explanatory faithfulness and localization reliability of saliency-driven methods, especially for transformer-based architectures (Panboonyuen, 19 Jan 2026). Future work is recommended to incorporate multi-institutional scans, richer DICOM metadata, cross-validation protocols, and linked clinical endpoints (Chaddad et al., 2 Jul 2025).

The IQ-OTH/NCCD dataset provides a valuable multiclass resource for algorithm development, but its generalizability and representativeness require further validation through expanded, demographically diverse cohorts and rigorous annotation workflows.