Classification of Benign and Malignant Renal Tumors Based on CT Scans and Clinical Data Using Machine Learning Methods

Jie Xu, Xing He, Wei Shao, Jiang Bian, Russell Terry

TL;DR

This paper introduces a machine learning framework to differentiate benign from malignant renal tumors using readily available clinical data and single-phase CT scans. By combining traditional ML and deep learning approaches with both clinical and radiomic features, the study achieved promising results in improving diagnostic accuracy, potentially reducing overtreatment of benign tumors.

What Problem Does This Paper Address?

Renal tumors present a diagnostic challenge, as a significant portion (up to 20%) of smaller masses (≤4 cm) are benign. However, accurately predicting malignancy pre-surgery using imaging alone is difficult, leading to concerns about overtreatment. Current methods often rely on multi-phase CT scans, which involve higher radiation exposure, longer scanning times, and increased costs. This study aims to develop a more efficient and accessible method for pre-operative classification by leveraging single-phase CT scans and clinical data.

Methods

The researchers explored a combination of traditional machine learning (ML) algorithms and deep learning (DL) models.

  • Traditional ML: XGBoost and Random Forest (RF) were employed. These models were trained using three configurations: clinical attributes only, radiomic features only, and a combination of both.
  • Deep Learning: A Multilayer Perceptron (MLP) was used for clinical and radiomic features, while a 3D Convolutional Neural Network (3DCNN) was developed to process volumetric CT scan data.
  • Feature Extraction: Radiomic features were extracted from CT scans using the PyRadiomics platform, encompassing various categories like first-order statistics, shape-based features, and texture features (e.g., gray level co-occurrence matrix). Clinical attributes included demographics, vital signs, and comorbidities.
  • Model Training and Evaluation: An 80/20 train-test split was used. ML models were optimized using five-fold cross-validation and randomized search. Performance was evaluated using AUC, precision, recall, and specificity, with bootstrapping for confidence intervals. Kruskal-Wallis and Mann-Whitney U tests were used for statistical comparisons.

Datasets

The study utilized data from the KiTS21 cohort, comprising 300 patients who underwent partial or radical nephrectomy for radiographically detected renal tumors between 2010 and 2020. This dataset included clinical attributes and preoperative single-phase CT scans. The data is publicly available on GitHub.

Key Insights and Ideas

  • Synergy of Data Modalities: The most significant finding is that combining clinical attributes with radiomic features extracted from CT scans yields superior performance compared to using either data source alone. This highlights the value of a multimodal approach.
  • Promise of Single-Phase CT: The study demonstrates that even with single-phase CT scans, which are more accessible and less burdensome than multi-phase scans, accurate classification is achievable, offering a practical advantage.
  • Traditional ML vs. Deep Learning: While both ML and DL models showed promise, the traditional ML model (RF) using combined features slightly outperformed the DL model in precision. This could be attributed to the relatively small dataset size, which might limit the DL model's ability to learn complex patterns effectively.
  • Clinical Relevance: The ability to accurately classify renal tumors pre-operatively can significantly impact clinical decision-making, potentially reducing unnecessary surgeries for benign lesions and improving patient outcomes.

Key Results

The combination of clinical and radiomic features, when used with traditional ML models, achieved the best overall performance. Specifically, this approach yielded:

  • AUC: 0.719 (95% CI: 0.712–0.726)
  • Precision: 0.976 (95% CI: 0.975–0.978)
  • Recall: 0.683 (95% CI: 0.675–0.691)
  • Specificity: 0.827 (95% CI: 0.817–0.837)

The study also noted that malignant tumors were more prevalent in male patients, aligning with previous observations.

Limitations and Future Work

The authors acknowledge several limitations:

  • Dataset Size and Imbalance: The study used a relatively small dataset (300 patients) with a significant imbalance, with a high prevalence of malignant tumors (91.67%). This imbalance could affect model performance, especially for benign tumors.
  • Limited Clinical Attributes: Due to the dataset's origin (a segmentation challenge), the available clinical attributes were limited.
  • Limited Model Exploration: Only a few ML and DL methods were tested.
  • Exclusion of Rare Tumor Types: Tumor types recently added to the WHO classification were not included.

Future work should focus on:

  • External Validation: Validating the proposed model and features on independent datasets.
  • Larger and More Diverse Datasets: Incorporating more data, particularly for benign tumors, and exploring a wider range of tumor subtypes.
  • Integration of Additional Data: Including clinical notes, genetic data, and other supplementary sources for enhanced classification.
  • Exploring Advanced ML Techniques: Investigating a broader spectrum of ML algorithms.
  • Identifying Specific Tumor Subtypes: Developing models capable of differentiating between benign and malignant lesions and identifying specific tumor subtypes.

Summary updated

Abstract

Up to 20% of renal masses ≤4 cm is found to be benign at the time of surgical excision, raising concern for overtreatment. However, the risk of malignancy is currently unable to be accurately predicted prior to surgery using imaging alone. The objective of this study is to propose a machine learning (ML) framework for pre-operative renal tumor classification using readily available clinical and CT imaging data. We tested both traditional ML methods (i.e., XGBoost, random forest (RF)) and deep learning (DL) methods (i.e., multilayer perceptron (MLP), 3D convolutional neural network (3DCNN)) to build the classification model. We discovered that the combination of clinical and radiomics features produced the best results (i.e., AUC [95% CI] of 0.719 [0.712–0.726], a precision [95% CI] of 0.976 [0.975–0.978], a recall [95% CI] of 0.683 [0.675–0.691], and a specificity [95% CI] of 0.827 [0.817–0.837]). Our analysis revealed that employing ML models with CT scans and clinical data holds promise for classifying the risk of renal malignancy. Future work should focus on externally validating the proposed model and features to better support clinical decision-making in renal cancer diagnosis.
BibTeX
@article{xu2023classification,
  author = {Jie Xu and Xing He and Wei Shao and Jiang Bian and Russell Terry},
  title = {{Classification of Benign and Malignant Renal Tumors Based on CT Scans and Clinical Data Using Machine Learning Methods}},
  journal = {Informatics},
  year = {2023},
  doi = {10.3390/informatics10030055},
}
Figure