Impact of slice thickness on reproducibility of CT radiomic features of lung tumors

Introduction

Radiomics is a new field that seeks to improve the physician’s visual perception of medical images with addition of more quantitative objectivity. The quantitative attributes from radiographic images are utilized to characterize spatial and textural patterns of lesions which can provide information about the heterogeneity associated with biological processes. Radiomics is a rapidly evolving field particularly in oncology to improve patient care, aid in treatment decision making, characterization, response to therapy and prognosis.^1–5

Lung cancer/carcinoma (LC) remains one among the most prevalent and familiar types of cancer that results in mortality notwithstanding recent improvements in healthcare. As, most detected LC are in the middle to late phase of the disease progression and have few management options left, hence, people with lung cancer have a 10-20% survival rate at 5 years following the diagnosis in most of the developed nations.^6,7 Radiomics and Machine learning methods have been used for classification of histological subtypes of LC, prediction of LC staging and outcome, response to treatment, prognosis of lung cancer.^8–11

Radiomics, a rapidly evolving field, employs quantitative attributes from medical images to enhance physician’s interpretation, particularly in oncology. Radiomics and machine learning models developed based on radiomic features play crucial roles in classifying histological subtypes lung cancer. Evaluating the variability of radiomic features (RF) is important as diagnosis and treatment decision made using these quantitative should be precise and reproducible. Recent studies have shown that the CT technical parameters such as exposure factors, slice thickness (ST) and image reconstruction algorithms (IRA) can significantly affect the values of RF. Experts have recommended that for training predictive models using radiomics based machine learning models, only reproducible RF should be considered.^12–14 The reproducibility of texture analysis of lung tumors is unclear and there is scarcity of research that has delved into how ST affects variability of RF in lung tumors. The research gap of the study centers on the application of Radiomics in LC, emphasizing the need for a deeper understanding of the reproducibility of RF in the context of lung tumors. Hence, aim of the study is to evaluate the influence of ST on reproducibility of CT-RF for lung tumors.

Methods

This is a prospective study. The study was commenced upon approval from the Institutional Ethical committee of Kasturba Medical College and Hospital, Manipal, India on 12^th August 2022 (IEC:193/2022) followed by the enrolment of the first subject after registration in the Clinical Trial Registry – India (CTRI) registration (CTRI/2022/09/045554) on 15^th September 2022, and continued till 30^th April 2023.

Segmentation

The Digital Imaging and Communication in Medicine (DICOM) CECT sections of two different slice thickness (2 mm and 5 mm) of the same patient were loaded into 3D slicer (version 4.10.2) and a radiologist (Bharath J L) with over 10 years of experience manually delineated the tumours (see Figure 1 for an example). The segmentation was performed using lung window (Window Width (WW): 1500 HU and Window Level (WL): -600 HU). All pulmonary nodules/lesions present in the right, left, and bilateral lungs were segmented rather than solely selecting the largest or most prominent nodule/lesion. The segmentation of the nodule was performed while excluding the airways, blood vessels, or bronchi. We extracted RF from the segmented regions of the lung nodules using both 2-mm and 5-mm ST.

Figure 1. An axial CT image of a 52-year-old male with adenocarcinoma showing manual segmentation of tumour using 3D Slicer at (a) 2mm and (b) 5mm slice thickness.

Results

A total of 32 cases (18 males and 14 females) with LC [Non-Small Cell Lung Cancer (NSCLC) – 71.8%, Small Cell Lung Cancer (SCLC) – 28.1%)] with mean age of were included 53.16 ± 10.25.

A total of 3424 Radiomic Features (RF) measurements (107 RF per study) were extracted. Among them 66 (61.6%) RF exhibited significant difference between two the slice thickness measurements, while 41 (38.3%) RF did not show significant difference between the two slice thickness measurements (Figure 2; Table 3).

Figure 2. Number of significant (p<0.05) and non-significant (p>0.05) features in each RF category between 2-mm and 5-mm slice thickness.

Reproducibility of RF

It was found that out of 14 shape-based features 8 (57.1%), out of 14 Gray Level Dependence Matrix (GLDM) RF 5 (35.71%), out of 24 for Gray Level Co-occurrence Matrix (GLCM) RF 3 (12.5%), out of 18 first order RF 8 (44.4%), out of 16 Gray level run length matrix (GLRLM) RF 4 (25%), out of 16 Gray level size zone matrix (GLSZM) RF 1 (6.25%) were found to be reproducible. All 5 neighboring gray tone difference matrix (NGTDM) RF were found to be not reproducible. Among the seven features categories, the shape-based features (57.1%) showed the maximum reproducibility whereas NGTDM based features showed negligible reproducibility (Table 4). The mean CCC of RF categories were shown in Figure 3.

Table 4. Percentage of reproducibility of RF between 2-mm and 5-mm ST for each category.

RF Category	Excellent n(%)	Good n(%)	Moderate n(%)	Weak n(%)
Shape (n=14)	2 (14.2%)	3 (21.4%)	3 (21.4%)	6 (42.8%)
GLDM (n=14)	-	-	5 (35.7%)	9 (64.2 %)
GLCM (n=24)	-	-	3 (12.5%)	21 (87.5%)
First order (n=18)	-	1 (5.5%)	7 (38.8%)	10 (55.5%)
GLRLM (n=16)	-	1 (6.25 %)	3 (18.75 %)	12 (75 %)
GLSZM (n=16)	-	-	1 (6.25%)	15 (93.75%)
NGTDM (n=5)	-	-	-	5 (100%)

Figure 3. Mean concordance correlation coefficient (CCC) of each radiomic feature category between 2-mm and 5-mm slice thickness.

Discussion

In the present study, we assessed the impact of slice thickness on the reproducibility of CT radiomic features (RF) for lung tumors. Few previous studies had addressed the influence of exposure parameters such as tube voltage (kV_P), tube current (mA), image reconstruction algorithms (IRA), CT Scanner vendors on RF in CT for conditions like liver fibrosis, metastatic liver lesions, pancreatic neuroendocrine neoplasm.^15–18 Variability of acquisition parameters could affect the diagnostic performance of radiomic signatures in oncologic patients.^18,19 Limited studies had investigated the impact of ST on reproducibility of CT-RF in lung tumors.

In this study, the category of shape-based RF (57.1%) exhibited the highest reproducibility compared to other RF categories. These shape based features demonstrated robustness due to presence of low-frequency components and the reliance on segmented boundaries resulting in consistent reproducibility across changes in ST. Findings by Erdal et al.²⁰ & Lu et al.²¹ supported this, revealing that RF describing tumor dimension, shape of boundaries, low-order density frequencies, and rough features were less sensitive to image setting parameters, in contrast to features characterizing sharpness of boundaries, high-order density frequencies and smooth features. Both studies analyzed the combination of ST with IRA (lung and standard) and noted that shape-based features were less effected by change in slice thickness and reconstruction algorithm. They also observed that the thinner slices with sharper reconstructions had fewer reproducible features compared to thicker slices with smoother reconstructions.

The GLDM category features in our study, such as HGLE, DE, SDE, DNU, LDHGLE demonstrated moderate reproducibility. A study by Emaminejad et al.²² in non-contrast chest CT (NCCT) identified that GLDM DE, DNU, GLNU were reproducible against the dose and kernel variations with varying slice thickness. Unlike our study, none of the previous research mentioned the reproducibility of GLDM features concerning slice thickness alone.

Within the GLCM category in our study, only two features showed reproducibility with variations in slice thickness. Similar results were documented by Erdal et al.²⁰ & Kim et al.²³ indicating that GLCM category (19.4 % & 25 %) had lower reproducibility compared to other RF categories. We observed that first-order features (44.4%) had the second highest reproducibility. Studies by Erdal et al.,²⁰ Park s et al.,²⁴ Choe J et al.²⁵ reported that first-order features exhibited the most reproducibility across various imaging parameters. Park s et al.²⁴ and Choe J et al.²⁵ reported that convolution network-based super resolution (SR) algorithms and kernel-converted images had reduced effects on the reproducibility of RF with variations in slice thickness and reconstruction kernels. Yang et al.²⁶ employed a resampling technique to standardize the voxel measurement of both thick and thin section CT images to 1x1x1 mm³ using linear interpolation and observed that, following resampling of thicker images, 202 RF (66.2%, 202/305) exhibited a noteworthy reduction in variability of RF compared to the original non-resampled data (Table 5).

For the GLRLM and GLSZM categories, reproducibility rates were 25% and 6.25 %, respectively, in the current study. A Study by Emaminejad et al.²² similarly found that GLRLM Run length non uniformity (1 of 9 features) and GLSZM (1 of 10 features) displayed very limited reproducibility against the dose and kernel variations with varying slice thickness. Contrary to the study reported by Liu J et al.²⁷ which demonstrated that NGTDM exhibited good reproducibility, we did not observe any reproducible features in NGTDM. The reason for this disparity is attributed to differences in technical parameters, specifically in terms of dose variation, rather than slice thickness.

The study has few limitations. Firstly, the sample size was relatively small, as it is time bound study with prospective data collection of patients who underwent CT scan with histopathological proven cases of lung cancer. A larger sample size is required to confirm the reproducibility of RF with slice thickness. This may limit the generalizability of the findings to a broader population. Secondly, we did not analyze whether a thinner slice thickness would result in better performance of radiomic models for predicting lung cancer. Thirdly, a single image acquisition variable such as slice thickness was examined to determine how it affects the reproducibility of radiomic features. Additionally, the study focused on a specific CT scanner model (128-Incisive CT by Philips Health care), and the results might be informed by scanner-specific characteristics. Generalizing the findings to other CT scanners would require further investigation.

In terms of potential implications for clinical practice, the study underscores the importance of considering the slice thickness when utilizing radiomics in the assessment of lung tumors. Clinicians should be aware that variations in slice thickness can introduce significant variability in RF. This finding emphasizes the need for standardizing imaging protocols, particularly in the context of lung cancer diagnosis and treatment planning. The results also highlight the critical role of shape-based features in radiomics, as they demonstrated the highest reproducibility in this study. Clinicians incorporating radiomic analysis into their practice should be attentive to the choice of features, giving preference to those with higher reproducibility for more reliable and consistent results.

Conclusion

Radiomics has the potential to transform lung cancer diagnosis, follow-up, and therapy planning by enabling individualised management in a non-invasive and an economical manner. Our study found that ST is the main parameter impacting the reproducibility of CT-RF for lung tumours. The original contribution of this study lies in its systematic examination of the influence of ST on the reproducibility of CT-RF in lung tumors. By identifying specific categories of RF that are more or less affected by variations in ST, the study provides valuable insights for researchers and clinicians working with radiomics in the field of lung cancer. This information could contribute to the refinement of imaging protocols, the standardization of radiomic analyses, and the interpretation of radiomic data in oncology. The study also increases awareness regarding the significance of accurately configuring imaging acquisition parameters in the context of radiomic/radio genomic approaches. Standardization of technical parameters and protocols is necessary when conducting multicentre studies, as these factors can impact the diagnostic performance of Machine Learning (ML) models developed using radiomic features.