MRI‑based diffusion weighted imaging and diffusion kurtosis imaging grading of clear cell renal cell carcinoma using a deep learning classifier

Introduction

Clear cell renal cell carcinoma (ccRCC) is a malignant tumor arising from the renal epithelium, accounting for ~85% of RCC cases (1). Compared with other RCC subtypes, ccRCC is characterized by enhanced invasion capacity and poorer prognosis (2). The latest grading system, developed jointly by the World Health Organization (WHO) and International Society of Urology, categorizes ccRCC into four grades (I–IV). Grades I and II are considered low-grade tumors with a favorable prognosis, while grades III and IV are characterized as high-grade tumors associated with poor prognosis (3). Minimally invasive treatment approaches, such as nephron sparing surgery, ablation, and even active monitoring, are more suitable for the management of low-grade ccRCC, while high-grade ccRCC can require radical surgery (4). Currently, preoperative biopsy remains the gold standard for the pathological grading of ccRCC. However, biopsy is an invasive procedure that carries several risks, such as dissemination and sampling errors (5). Therefore, non-invasive and accurate grading of ccRCC using imaging techniques has emerged as a prominent research direction in recent years.

It has been reported that magnetic resonance imaging (MRI) can reveal several tumor characteristics, including tumor size, necrosis, bleeding, enhancement pattern, and venous thrombosis (4,6). However, MRI still lacks sufficient information to aid radiologists in distinguishing the pathological grade of ccRCC (7).

Radiomics is a science field that involves the extraction, analysis, and interpretation of quantitative imaging parameters. It utilizes computer-based post-processing technology to convert medical images into high-dimensional and quantitative imaging features. Utilizing model-building algorithms, these imaging features can be associated with tumor tissue pathology and heterogeneity, thus offering valuable insights for tumor classification and grading, gene localization, as well as early prediction of treatment response (4). Deep learning (DL) algorithms, a form of radiomics, automatically learn and extract intricate patterns and features from data via multi-level neural network models, thus enabling efficient data analysis, prediction, and decision-making (8). Currently, the association between diffusion weighted imaging (DWI) and diffusion kurtosis imaging (DKI) parameters and the pathological grade of ccRCC has been extensively studied (2,9–14). However, research on MRI-based DWI and DKI-based DL models for predicting the pathological grade of ccRCC remains limited.

Therefore, the present study aimed to validate a DL model based on MRI data for predicting the pathological grading of ccRCC prior surgery.

Materials and methods

Patients

The present study was approved by the Medical Ethics Committee of Zibo Central Hospital (approval no. 2017001), and all patients provided signed written informed consent prior to undergoing any examinations. The analysis included patients with histologically confirmed ccRCC, who were treated at Zibo Central Hospital between March 2017 and November 2021. A total of 108 patients with suspected renal tumors underwent MRI scan 1–2 weeks prior to surgery. The inclusion criteria were as follows: i) Pathologically confirmed ccRCC via postoperative histologic examination; ii) MRI protocol including both DWI and DKI; and iii) no prior history of radiotherapy, chemotherapy, or percutaneous biopsy prior to renal MRI scanning. The exclusion criteria were as follows: i) Pathologically confirmed non-ccRCC (such as papillary RCC, collecting duct carcinoma, angioleiomyolipoma, etc.); ii) Renal MRI protocol lacking DWI or DKI, with incomplete imaging data; iii) A history of radiotherapy, chemotherapy, or percutaneous biopsy prior to renal MRI scanning; iv) Poor quality of MRI images with obvious artifacts that interfere with the observation and analysis of lesions; v) ccRCC lesions that are too small, in special locations, or with other characteristics that make accurate imaging classification or parameter measurement impossible. Finally, a total of 79 patients with ccRCC were enrolled in the current prospective study. In addition, a total of 19 patients were excluded from the study due to non-ccRCC pathological types, including papillary RCC (n=11), collecting duct carcinoma (n=3), and angioleiomyolipoma (n=5). Among patients diagnosed with ccRCC, three were excluded because the solid components of the lesions were difficult to identify, four were excluded due to poor MRI image quality or significant artifacts, and three were excluded since the lesion size was <1 cm in length. The inclusion and exclusion criteria are listed in Fig. 1.

Figure 1. Pathway for patient recruitment. ccRCC, clear cell renal cell carcinoma.

MRI acquisition

All MRI examinations were carried using a 3-Tesla (T) scanner (Signa™ HDx; GE Healthcare) and a commercial 16-channel torso phased-array coil positioned at the renal level. All patients underwent respiratory training prior to MRI examinations and were scanned to the supine, feet-first position. Each patient was subjected to two MRI sequence types, namely DWI and DKI. DWI was carried out using a single-shot echo planar imaging sequence in the axial plane, also incorporating imaging technology, and respiratory-triggered acquisition, with b value of 0,800 s/mm2, a field of view (FOV) of 410 mm, and repetition time (TR)/echo time (TE) of 5,400/64 ms. The layer thickness was 5.0 mm with 1.0-mm overlap, while a total of 19 slices were acquired. Finally, 128×128 matrix images were produced, with a total collection time of 1 min. For DKI, a single-shot echo planar imaging sequence in the axial plane with respiratory-triggered acquisition was performed. DKI images were captured with 30 diffusion gradient directions with three different b values for each direction (0, 1,000, and 2,000 s/mm2). The remaining MRI parameters were as follows: a FOV of 410 mm, TR/TE of 5,500/85 ms, and slice thickness of 5.0 mm with a 1.0-mm overlap. The image matrix was 128×128 with a collection time of approximately 3 min and 7 sec. The center level of the DKI was the same as that of DWI. Finally, the raw data from the DWI and DKI were transferred and processed using Functool software on the Advantage Workstation (release 4.5). For DWI, apparent diffusion coefficient (ADC) maps were generated, while for DKI, axial kurtosis (Ka), fractional anisotropy (FA), radial kurtosis (Kr), mean kurtosis (MK), and mean diffusivity (MD) maps were calculated.

Image processing

A total of 569 ADC images, including 385 in the training set and 184 in the test set, were acquired. The program effectively processed all 569 ADC images. For the DKI/FA scans, a total of 560 images, 363 in the training set and 197 in the test set, were captured, which were all successfully read by the program. In addition, for the DKI/Ka scans, there were a total of 540 effectively read images, with 351 in the training set and 189 in the test set. The DKI/Kr scans totaled 573 images, comprising 365 and 208 in the training and test sets, respectively. All images were also successfully read by the program. Similarly, the DKI/MD scans consisted of 555 effectively read images (363 in the training set and 192 in the test set). For the DKI/MK scans, 558 images, 361 in the training set and 197 in the test set, were included, which were all also successfully accessed. Finally, the DWI scans consisted of 546 images, with 377 in the training set and 169 in the test set, all of which were effectively read. The potential reasons for inconsistencies among the different parameter maps are illustrated in Fig. 2.

Figure 2. Flow chart for MRI images. ADC, apparent diffusion coefficient; DKI, diffusion kurtosis imaging; DWI, diffusion weighted imaging; FA, fractional anisotropy; Ka, axial kurtosis; Kr, radial kurtosis; MD, mean diffusivity; MK, mean kurtosis.

Data processing

During the collection of DWI mono-exponential mode images, it was noticed that ccRCC lesions were primarily observed around the central slice (Fig. 3A), while the remaining slices provided limited diagnostic value. Therefore, the appropriate slices, from where experts could visually identify lesions, were manually selected. The filtering process was carried out by two professional radiologists. More particularly, the one radiologist filtered all images, and the second reviewed the filtering results. In case of possible disagreement, the final decision was made through consultation. In addition, frequency distribution maps of the ADC and DWI sequences indicated that although the intensity of mono-exponential mode images varied, the high-frequency intensities were mainly concentrated within the low-value range (Fig. 3B). To normalize image intensity among different exponential imaging modes, the high-frequency range of the ADC and DWI sequences was scaled to [0, 255]. More particularly, the effective intensity ranges of the ADC and DWI sequences were [-256,500] and [0,120], respectively. The DKI-derived sequences remained unchanged.

Figure 3. Overview of data processing. (A) Data collection, (B) image preprocessing, and (C) model training and test. ADC, apparent diffusion coefficient; DKI, diffusion kurtosis imaging; DWI, diffusion weighted imaging; FA, fractional anisotropy; Ka, axial kurtosis; Kr, radial kurtosis; MD, mean diffusivity; MK, mean kurtosis.

Model training and test

A supervised DL model was used to predict the pathological grades of ccRCC based on the aforementioned data. As shown in Fig. 3C, the training set was used to optimize the model parameters and to update the model weights. However, the optimal model weights were retained. During the testing phase, model weights were accordingly loaded into the model architecture to predict the lesion grades of the test samples. The detailed model architecture is depicted in Fig. 4. Due to the limited data, the VGG-16 model was selected as the backbone architecture. The model comprised of five convolutions and three linear classification layers. The first two convolutional blocks consisted of two convolutional modules, each including a convolution layer, a batch normalization (BatchNorm) layer and a ReLU function module, and a max-pooling layer. The remaining three convolutional layers consisted of three convolutional modules and a max-pooling layer. The number of convolutional channels were 64, 128, 256, 512, and 512 respectively. The classification layers primarily included a linear-mapping layer, a BatchNorm module and a ReLU activation module. In the present study, patients diagnosed with grade I and II tumors were regarded as low-grade patients, marked as 0, while those with grade III and IV tumors were classified as high-grade, marked as 1. Therefore, the grading of ccRCC lesions was considered as a binary classification problem. The final linear classification channel was set to 2, followed by a softmax classification.

Figure 4. Detailed model architecture. ADC, apparent diffusion coefficient; DKI, diffusion kurtosis imaging; DWI, diffusion weighted imaging; FA, fractional anisotropy; Ka, axial kurtosis; Kr, radial kurtosis; MD, mean diffusivity; MK, mean kurtosis; Conv, convolution; Conv2d, 2D convolution; MaxPool2d, 2D max-pooling.

Statistical analysis

Data analyses were performed using the IBM SPSS Statistics for Windows, version 22.0 (IBM Corp.). The Shapiro-Wilk test for normality and Levene's test for homogeneity of variances were carried out on the age variable. Both tests verified that the data met the assumptions of normality and homogeneity of variance. The differences in age between patients with low- and high-grade ccRCC were compared utilizing an independent samples (unpaired) t-test. The χ2 test was employed to compare the differences in sex between low- and high-grade tumors. For tumor side, Fisher's exact test was used instead, as the assumption of expected counts for the χ2 test was not met. P<0.05 was considered to indicate a statistically significant difference.

Results

Patient and tumor characteristics

Based on the pathological results, a total of 79 patients with ccRCC were enrolled, including 40 cases of low-grade disease (six cases of grade I and 34 cases of grade II) and 39 cases of high-grade disease (25 cases of grade III and 14 cases of grade IV). The baseline clinical characteristics of the included patients are listed in Table I. Therefore, no statistical differences were obtained between patients with low- and high-grade ccRCC.

Table I. Baseline clinical characteristics of the study cohort.

Table I.

Baseline clinical characteristics of the study cohort.

Characteristics	Low-grade (grade I and II; N=40)	High-grade (grade III and IV; N=39)	T/χ2	P-value
Mean age ± SD, years (range)	57.25±13.04 (36–80)	60.33±10.45 (27–79)	1.158	0.250
Sex, n (%)			1.764	0.241
Male	23 (57.5)	28 (71.8)
Female	17 (42.5)	11 (28.2)
Side, n (%)			1.143	0.736
Left	19 (47.5)	17 (43.6)
Right	20 (50.0)	22 (56.4)
Both sides	1 (2.5)	0 (0.0)
Grade, n (%)
I	6 (15.0)	-
II	34 (85.0)	-
III	-	25 (64.1)
IV	-	14 (35.9)

Model performance

The VGG-16 model was used for binary classification, and the particular experimental results are presented in Table II. Among the investigated imaging parameters, MK achieved the highest accuracy, followed by Kr and Ka. The precision, recall, F1-score, and accuracy values for MK were 81.48, 76.04, 74.08, and 76.04%, respectively. In addition, those for Kr were 75.51, 75.36, 75.42, and 75.36%, respectively. Finally, for Ka the corresponding rates for precision, recall, F1-score and accuracy were 81.39, 71.81, 68.31, and 71.81%, respectively.

Table II. Results of the binary classification experiment for clear cell renal cell carcinoma.

Table II.

Results of the binary classification experiment for clear cell renal cell carcinoma.

Imaging parameter	Precision, %	Recall, %	F1-score, %	Accuracy, %
ADC	69.96	69.32	65.99	69.32
DKI/FA	71.21	71.35	70.94	71.35
DKI/Ka	81.39	71.81	68.31	71.81
DKI/Kr	75.51	75.36	75.42	75.36
DKI/MD	78.13	71.74	65.36	71.74
DKI/MK	81.48	76.04	74.08	76.04
DWI	70.79	70.24	70.44	70.24

[i] ADC, apparent diffusion coefficient; DKI, diffusion kurtosis imaging; FA, fractional anisotropy; Ka, axial kurtosis; Kr, radial kurtosis; MD, mean diffusivity; MK, mean kurtosis; DWI, diffusion weighted imaging.

Discussion

The current study aimed to evaluate the effectiveness of different MRI sequences for a binary DL model that was established for the preoperative pathological grading of ccRCC. Among the MRI-based approaches applied, DKI/MK displayed the highest grading accuracy. The best-performing DL model showed strong diagnostic performance, with accuracy and precision rates of 76.0 and 81.5% respectively. These results indicated that the DL method was effective for ccRCC grading evaluation, thus offering an non-invasive and valuable approach for ccRCC grading, and providing a significant tool that could support clinicians in guiding individualized treatment strategies. In recent decades, several studies have investigated the association between DWI, DKI, and radiomics, and the pathological grading of ccRCC, thus supporting the increasing attention in this field.

DWI is a non-invasive method used to detect the diffusion movement of water molecules within living tissues. The net movement of water molecules can be quantified using the apparent ADC value (15). It has been reported that the ADC values of ccRCC tends to decrease as the pathological grading increases (2,11). DKI is an advanced functional MRI technology developed based on DWI, incorporating a non-Gaussian model. This method is based on the fact that water molecule diffusion in living tissues follows a non-Gaussian distribution due to the effect of different factors, such as cell membranes and intercellular organelles. DKI can provide a more accurate assessment of microstructure changes via calculating the deviation from the Gaussian distribution between real and ideal diffusion states (2,16). The kurtosis parameters of DKI include MK, Ka and Kr, while the diffusion parameters include MD, FA, axial diffusivity, and radial diffusivity (17).

MK represents the mean value of diffusion kurtosis values measured in each direction. It serves as an average index of the limitations experienced by water molecule diffusion within tissues, thereby indirectly reflecting the complexity of the tissue structure (12). Additionally, Kr and Ka are used to elucidate tissue diffusion patterns in distinct orientations, with Kr reflecting the mean kurtosis along the principal axis of the diffusion tensor, and Ka perpendicular to the principal axis. Both metrics are considered as indicative markers of tissue complexity (18).

However, there are only a few reports on the application of DKI in ccRCC classification. Wu et al (14) explored the potential of DKI in grading ccRCC among 91 patients. The results showed that MK could more accurately grade ccRCC. These findings were consistent with the results of the present study. In addition, Cao et al (2) evaluated 89 patients with confirmed ccRCC using DKI on a 3-T MRI scanner. Consistent with the results of the current study, the above study reported that MK exhibited the highest area under the curve (AUC).

In addition, Ye et al (13) utilized a 3.0-T MRI scanner to examine 148 patients with ccRCC, using the DKI approach with three b values (0, 1,000, and 2,000 s/mm2) and 30 diffusion directions. The results demonstrated that the AUC value for MD was the highest, thus suggesting that MD could be the most valuable parameter for grading ccRCC using DKI. Similarly, Zou et al (12) evaluated 108 patients with confirmed ccRCC with DKI (b values, 0, 1,000, and 2,000 s/mm2) utilizing a 3.0-T MRI scanner. The results of this study consistently showed that the MD values displayed the highest diagnostic performance for ccRCC grading. Additionally, in the study by Cheng et al (9), a total of 65 patients with pathologically confirmed ccRCC were assessed using DKI sequences with three b values (0, 500, 1,000 s/mm2), also using a 3.0-T MRI scanner. Therefore, this study also showed that MD values exhibit a good diagnostic efficiency for ccRCC grading. However, the results of the aforementioned studies were not consistent with those observed in the current study. This could be due to the following reasons: i) MK indirectly reflects the complexity of the tissue structure, while MD characterizes the overall diffusion of water molecules. In high-grade ccRCC, the presence of necrotic areas can lead to increased restriction on water molecule diffusion, potentially resulting in falsely enhanced MD values. This could obscure the actual microstructural differences resulting from cellular density and architectural disorders, thus contributing to the superior performance of MK compared with MD; ii) In the study by Cheng et al (9) a maximum b value of only 1,000 s/mm2 was employed, which could be suboptimal. Rosenkrantz et al (10) suggested that in order to ensure reliable results, the maximum b value for kidney DKI should not be less than 1,500 s/mm2; and iii) In the aforementioned studies, including the current one, the majority of patients with ccRCC undergone radical or partial nephrectomy. Therefore, the number of cases with high-grade and advanced ccRCC was relatively small, while the sample size between the two groups was quite different, thus introducing selection bias and affecting the comparisons between groups.

Radiomics, a significant scientific field, is associated with the extraction, analysis, and interpretation of quantitative imaging parameters. It utilizes computer post-processing technology to transform conventional imaging data into high-dimensional and quantitative imaging data. Core radiomic-related methods include texture analysis (TA), machine learning (ML), and DL (4). TA is a quantitative imaging technique that is used to extract additional information from the grayscale distribution of pixels or voxels within medical images, thus providing quantitative statistical parameters (19). ML and DL, subfields of artificial intelligence, primarily focus on developing algorithms that can learn from and improve via data analysis without requiring explicit programming in advance (20). ML models commonly use TA-derived parameters as input features, thus enhancing the sensitivity of medical imaging-mediated diagnosis (21). By contrast, DL models do not require manual feature selection, as the algorithm itself can autonomously determine which features are the most suitable for model development (22). As a form of radiomics, DL algorithm can automatically learn and extract complex patterns and features from data via using multi-level neural network models, thus enabling efficient data analysis, prediction and decision making (8). Although several studies have been conducted on the pathological grading of ccRCC using MRI-based approaches (7,23,24), to the best of our knowledge, no studies have yet employed the DKI-based approach.

In a previous study, Pan et al (24) extracted MRI texture features from 89 patients with ccRCC. Therefore, conventional sequences, such as T1-weighted imaging (T1WI), T2WI and contrast-enhanced T1WI (CE-T1WI) were classified as conventional MRI (cMRI), while Dixon-MRI, blood oxygen level-dependent MRI and SWI were classified as functional MRI (fMRI). The accuracy rates of cMRI and fMRI were 49.56 and 70.80%, respectively. However, their approach primarily relied on manually defined features, and therefore not all imaging data could be captured. By contrast, DL can artificially outline regions of interest (ROI), and automatically extract image features, thus compensating for the lack of TA (23). Chen et al (7) combined T1WI, T2WI, CE-T1WI, and DWI sequences to establish a ML model for the pathological grading of ccRCC, including 99 patients. The training and validation accuracies were 95.9 and 86.2%, respectively. However, ML depends on manual mapping of ROIs, which can result in manual errors in the boundary of the lesion. By contrast, DL can directly identify the characteristics of the lesion from the model. In addition, Zhao et al (23) used T1-enhanced and T2WI sequences to establish DL models using data from 430 patients with ccRCC. The reported accuracy rates for T1 enhanced, T2WI, and their combination were 71, 71, and 90%, respectively. In the present study, the accuracy of the established DL model, using the DKI/MK, DKI/Kr, DKI/Ka, DKI/FA, and DKI/MD parameters, was higher compared with that reported for the individual sequence model by Zhao et al (23).

Different pathological grades of ccRCC can be associated with different surgical approaches and prognoses (25,26). Currently, preoperative biopsy remains the gold standard for determining the pathological grade of ccRCC. However, biopsy is an invasive examination, which is associated with particular risk factors, such as tumor seeding and sampling error (5). By contrast, DL models can predict the pathological grade of ccRCC under non-invasive conditions, thus assisting clinicians in developing personalized treatment strategies for patients with ccRCC.

However, this study has several limitations. Firstly, the number of patients with ccRCC included in the study was relatively small. Therefore, a multicenter study with a larger patient cohort should be performed in the future. In addition, due to the limited number of patients with grade I and IV ccRCC, binary classification was employed. Future studies with more patients of this grade and a four-classification approach should be considered to improve the specificity of pathological grading.

Overall, the present study validated a DL-based model using DWI and DKI imaging to accurately and non-invasively distinguish between low- and high-grade ccRCC. Therefore, this model could assist clinicians in guiding individualized clinical treatment of patients with ccRCC under non-invasive conditions.

Acknowledgements

Not applicable.

Funding

Funding: No funding was received.

Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

Authors' contributions

JC and WZ conceived and designed the experiments. XL, YaL, CX, ZD and WZ performed the experiments. KL, YuL, XS, DK and PS analyzed the data. JC, CX and ZD contributed to the provision of materials and analysis tools. YuL, XS and WZ confirm the authenticity of all the raw data. All authors have read and approved the final version of the manuscript.

Ethics approval and consent to participate

The study protocol was approved by the Medical Ethics Committee of Zibo Central Hospital (approval no. 2017001; Zibo, China). Written informed consent was obtained from all patients prior to enrollment.

Patient consent for publication

All patients or their next of kin consented to publication of the article.

Competing interests

The authors declare that they have no competing interests.

References