Article published online: 2024-12-11

THIEME
Original Article 395

Deep Learning for Detecting and Subtyping

Renal Cell Carcinoma on Contrast-Enhanced CT
Scans Using 2D Neural Network with Feature
Consistency Techniques
Amit Gupta1 Rohan Raju Dhanakshirur2 Kshitiz Jain3 Sanil Garg1 Neel Yadav1 Amlesh Seth4
Chandan J. Das1

1 Department of Radiodiagnosis and Interventional Radiology, All Address for correspondence Chandan J. Das, MD, PhD, DNB, FRCP
India Institute of Medical Sciences, New Delhi, India Edin, Department of Radiodiagnosis and Interventional Radiology, All
2 Amarnath and Shashi Khosla School of Information Technology, India Institute of Medical Sciences, Ansari Nagar, New Delhi 110029,
Indian Institute of Technology Delhi, New Delhi, India India (e-mail: [email protected]).
3 Yardi School of Artificial Intelligence, Indian Institute of Technology
Delhi, New Delhi, India
4 Department of Urology, All India Institute of Medical Sciences, New
Delhi, India

Indian J Radiol Imaging 2025;35:395–401.

Abstract Objective The aim of this study was to explore an innovative approach for developing
deep learning (DL) algorithm for renal cell carcinoma (RCC) detection and subtyping on
computed tomography (CT): clear cell RCC (ccRCC) versus non-ccRCC using two-
dimensional (2D) neural network architecture and feature consistency modules.
Materials and Methods This retrospective study included baseline CT scans from 196
histopathologically proven RCC patients: 143 ccRCCs and 53 non-ccRCCs. Manual
tumor annotations were performed on axial slices of corticomedullary phase images,
serving as ground truth. After image preprocessing, the dataset was divided into
training, validation, and testing subsets. The study tested multiple 2D DL architectures,
with the FocalNet-DINO demonstrating highest effectiveness in detecting and classify-
ing RCC. The study further incorporated spatial and class consistency modules to
enhance prediction accuracy. Models’ performance was evaluated using free-response
receiver operating characteristic curves, recall rates, specificity, accuracy, F1 scores,
and area under the curve (AUC) scores.
Keywords Results The FocalNet-DINO architecture achieved the highest recall rate of 0.823 at
► renal cell carcinoma 0.025 false positives per image (FPI) for RCC detection. The integration of spatial and
► deep learning class consistency modules into the architecture led to 0.2% increase in recall rate at
► classification 0.025 FPI, along with improvements of 0.1% in both accuracy and AUC scores for RCC
► detection classification. These enhancements allowed detection of cancer in an additional 21
► F1 score slices and reduced false positives in 126 slices.

article published online DOI https://doi.org/ © 2024. Indian Radiological Association. All rights reserved.
December 11, 2024 10.1055/s-0044-1800804. This is an open access article published by Thieme under the terms of the
ISSN 0971-3026. Creative Commons Attribution-NonDerivative-NonCommercial-License,
permitting copying and reproduction so long as the original work is given
appropriate credit. Contents may not be used for commercial purposes, or
adapted, remixed, transformed or built upon. (https://creativecommons.org/
licenses/by-nc-nd/4.0/)
Thieme Medical and Scientific Publishers Pvt. Ltd., A-12, 2nd Floor,
Sector 2, Noida-201301 UP, India
396 Predicting Renal Cell Carcinoma Subtypes and Grading Using 2D Neural Networks Gupta et al.

Conclusion This study demonstrates high performance for RCC detection and
classification using DL algorithm leveraging 2D neural networks and spatial and class
consistency modules, to offer a novel, computationally simpler, and accurate DL
approach to RCC characterization.

Introduction
and November 2021. Using these pathology reports as the gold
Renal cell carcinoma (RCC) represents a significant portion of standard, RCC patients were categorized into ccRCC and non-
cancer in adults, manifesting in a variety of histological ccRCC (including papillary and chromophobe RCCs). Abdomi-
patterns.1 Of these, the most prevalent and aggressive kind nal CT scans corresponding to these patients were obtained
is clear cell RCC (ccRCC).2 It is of paramount importance to from the hospital’s Picture Archiving and Communication
distinguish among RCC subtypes and determine their grades System (PACS) using their unique hospital identification num-
for patient prognosis and to customize treatment employing bers. Patients without a preoperative baseline scan in the PACS
the latest therapeutic options, including inhibitors of tyro- were excluded from the study. Thus, 196 RCC patients who had
sine kinase and vascular endothelial growth factor.3–5 Even in preoperative abdominal CT at our facility and thereafter
nonmetastatic RCC cases where surgery is the standard surgery were included in the analysis. In this cohort, 143
irrespective of the type of RCC, preoperative subtype classi- were ccRCCs and 53 were non-ccRCCs (42 papillary RCCs
fication still holds clinical relevance as it can guide surgical and 11 chromophobe RCCs). The authors confirm the avail-
planning, perioperative care, and risk stratification, especial- ability of, and access to, all original data reported in this study
ly for aggressive subtypes like ccRCC. However, current
diagnostic methods, including renal mass biopsies, are inva- Computed Tomography Protocol
sive and can lead to erratic results.6,7 The use of contrast A dual-source dual-energy 2
128 section multidetector CT
enhancement patterns in multiphasic computed tomogra- scanner (Somatom Definition Flash, Siemens Healthineers,
phy (CT) and magnetic resonance imaging (MRI) for RCC Germany) was used for multiphase CT imaging. A noncon-
characterization is also vulnerable to subjectivity and over- trast CT (NCCT) scan is the first step in the standardized renal
lapping imaging results.8,9 Consequently, there is a critical mass protocol used in our department. Corticomedullary
demand for dependable, noninvasive imaging biomarkers (CM) and nephrographic phase images are then obtained at
that can accurately classify and grade RCC, thereby improving 25 to 30 and 60 to 70 seconds, respectively, following intra-
diagnostic precision and patient management. venous injection of 100 mL of iodinated contrast (Omnipa-
Recent advances in deep neural networks (DNN) offer que 350, GE Healthcare) through a peripheral line at 3 to
promising avenues for enhancing RCC detection and classifi- 5 mL/s. Delayed excretory phase images are obtained after 4
cation through cross-sectional imaging.10–13 Previous re- to 5 minutes in patients with suspicion of renal collecting
search has shown that DNN models can accurately detect system involvement. The retrieved CT scans were made
and subtype RCC, highlighting their potential role in complex anonymous by removing the Digital Imaging and Commu-
medical image processing and patient prognostica- nications in Medicine (DICOM) metadata and reassigning
tion.12,14–16 However, the landscape of RCC characterization them with a new unique ID number for the study. The CM
and its management implications continues to evolve, ne- phase images were selected for model training and testing
cessitating further research. Challenges such as interobserv- due to their high sensitivity for RCC detection and character-
er variability in radiological interpretations and the ization per established CT protocols, with other contrast
overlapping image features of RCC subtypes call for more phases less consistently available due to protocol variability
refined and reliable diagnostic tools with feasibility for easy in the retrospective data.
integration into radiological practice.
With this background, our study aimed to investigate a Manual Tumor Annotation
novel method of training and constructing a deep learning The CM phase images from the retrieved CT abdomen scans
(DL) algorithm for RCC detection and subtyping (ccRCC vs. were converted into axial slices with a uniform resolution of
non-ccRCC) utilizing two-dimensional (2D) neural network 0.7 frames/mm and saved as individual Joint Photographic
architectures and feature consistency modules. Experts Group (JPEG) files for each patient. One radiologist
(with 9 years of body imaging experience) manually anno-
tated the tumors on these images. Each axial slice was
Materials and Methods
individually examined by the radiologist for the presence
CT Dataset Preparation of renal tumor. Using the freely accessible open-source
The Institute Ethics Committee approved this retrospective “LabelMe” software, manual tumor marking was done by
study. We reviewed the pathology report database of our the radiologist using rectangular bounding box placement.
tertiary care hospital for patients with histopathologically The manual labels were subsequently reviewed for correct-
proven RCC in nephrectomy specimens, between January 2016 ness by another radiologist (with 20 years of experience in

Indian Journal of Radiology and Imaging Vol. 35 No. 3/2025 © 2024. Indian Radiological Association. All rights reserved.
 Predicting Renal Cell Carcinoma Subtypes and Grading Using 2D Neural Networks Gupta et al. 397

Fig. 1 Architecture of the FocalNet-DINO-based object detection algorithm. CNN, convolutional neural network.

genitourinary imaging). This manual bounding box place- detection methods for this method. The FocalNet-DINO
ment by the radiologist was used as the ground truth for RCC design was found to be the most efficient of these. ►Fig. 1
detection during the training and testing of the networks. depicts the architecture of the FocalNet-DINO-based object
detection system. This intricate architecture involved divid-
Image Preprocessing and Data Stratification ing the input images into smaller patches, which were then
The collected image dataset underwent preprocessing to processed through a sophisticated ResNet-based backbone to
ensure consistency and quality. This included standardizing generate multiscale feature maps. Each feature map was
image dimensions by rescaling it to 224
224, normalizing refined using a series of encoder blocks equipped with
pixel intensities between 0 and 1 by employing the mean- focal-modulation-based self-attention mechanisms. This en-
max normalization algorithm, and applying standardized abled efficient and detailed interaction between the features
augmentation techniques such as random rotations, flips, of each patch, resulting in improved refinement. The en-
translations, changes in brightness, and contrast for en- hanced multiscale feature map generated by the encoder
hanced diversity. Following preprocessing, the data were layer was subjected to a query selection framework that
partitioned into three cohorts: training, validation, and employed a gated aggregation mechanism. This framework
testing. The training cohort, comprising 80% of the data, was useful in creating prospective tumor locations, which
was utilized to train the model. A separate validation cohort, were then refined during the decoder step. The decoder
constituting 10% of the data, facilitated fine-tuning the incorporated a deformable cross-attention module, a key
model parameters and guarded against overfitting. The component that focused on the feature maps corresponding
remaining 10% formed the testing cohort, reserved for to every query, facilitating accurate detection and classifica-
assessing the model’s performance on unseen data. Care tion of the tumor regions. The classification for each axial
was taken to maintain a balanced distribution of tumor section of the CT scan was determined based on the class
subtypes and to ensure that the data pertaining to one label of the most confident bounding box in that slice.
patient was kept in a single cohort to avoid label leaking.
Thus, 156 scans (115 ccRCCs and 41 non-ccRCCs) were used Incorporating Spatial and Class Consistency Modules
for training, while validation and testing included 20 scans in the DL Algorithm
each (14 ccRCCs and 6 non-ccRCCs per cohort). A novel component of our study was the 2D network’s
independent inference on each axial section of the CT scan.
Working of Deep Neural Network Architecture To refine the predictions, we took advantage of the intrinsic
In this study, we approached detecting and classifying RCC on spatial consistency across axial sections. To ensure spatial
CT as an image-based object detection problem. We experi- coherence, we altered the confidence levels of our predic-
mented with several 2D deep neural network–based object tions based on the spatial overlaps of three subsequent

Indian Journal of Radiology and Imaging Vol. 35 No. 3/2025 © 2024. Indian Radiological Association. All rights reserved.
398 Predicting Renal Cell Carcinoma Subtypes and Grading Using 2D Neural Networks Gupta et al.

Fig. 2 Overall pipeline of the proposed renal cell carcinoma (RCC) detection and classification model. CT, computed tomography.

frames. Furthermore, the mode of class labels across these a male-to-female ratio of 31:22 and an average tumor size of
frames was used to calculate each prediction’s class label. 6.4 cm (standard deviation:  2.3 cm).
Finally, the axial slices with annotated bounding boxes and
classification labels were combined into DICOM files for the Renal Cell Carcinoma Detection
final result. A schematic of the suggested RCC detection and We compared the performance of the proposed algorithm
classification model’s pipeline is presented in ►Fig. 2. against the ground truth for its detection performance. A
predicted bounding box was considered to be valid if the
Testing of the Networks and Data Analysis prediction confidence was greater than 0.5. We measured
We used the free-response receiver operating characteristic the recall rate of the proposed algorithm against various
(FROC) curve and recall rates at varying false-positive toler- tolerance levels of false positives per image (FPI) and the
ances to evaluate the performance of various network archi- results of the same can be observed in ►Table 1. Similarly,
tectures for RCC detection. The FROC curve offers a detailed the corresponding FROC curves can be found in ►Fig. 3.
graphical representation of sensitivity/recall values against The FocalNet-DINO architecture achieved a recall rate of
false-positives per image (FPI). Our criterion for a true 0.823 with 0.025 FPI, outperforming the next best-per-
positive prediction was based on the positioning of the forming network, DN-DEF, by 0.9%. The addition of spatial
predicted bounding box’s center within the ground truth and class consistency modules to our proposed design
bounding box, aligning with the medico-vision community improved the performance even more, with a 0.2% gain
standards. For the classification task, we assessed the perfor- in recall rate at 0.025 FPI. These results, particularly the
mance of the DL models using key metrics such as accuracy, recall rate of 0.825 at 0.025 FPI, provide strong evidence of
F1 scores, precision, recall, and area under the curve (AUC) the model’s high accuracy in RCC detection, underscored
scores. by the increased sensitivity achieved through spatial and
class consistency modules. ►Fig. 4 shows visualizations of
the ground truth labels and predictions by various net-
Results
works for different patients. The proposed feature consis-
Patient Demography and Tumor Characteristics tency strategies enabled the detection of cancer in an extra
The mean age of the 143 patients who had ccRCCs was 67.5 21 slices of the test dataset, increasing the model’s sensi-
years (range: 51–72 years), and the average size of their tumor tivity. In addition, the spatial and class consistency module
was 6.5 cm (standard deviation:  2.4 cm). There were 81 men demonstrated its effectiveness in improving prediction
and 62 women in the group. The average age of patient with accuracy by reducing the number of false positives in
non-ccRCCs (n ¼ 53) was 64.7 years (range: 48–70 years), with 126 slices.

Indian Journal of Radiology and Imaging Vol. 35 No. 3/2025 © 2024. Indian Radiological Association. All rights reserved.
 Predicting Renal Cell Carcinoma Subtypes and Grading Using 2D Neural Networks Gupta et al. 399

Table 1 Comparison of performance of various deep neural network–based object detection architectures for RCC detection and
classification

Model name Venue Detection results Classification results

[email protected] [email protected] [email protected] [email protected] [email protected] R@1 R@5 Acc F1 AUC
score score
DAB-DETR ICLR’22 0.763 0.845 0.884 0.915 0.927 0.937 0.969 0.942 0.874 0.976
DAB DEF. ICLR’22 0.786 0.849 0.878 0.927 0.942 0.956 0.977 0.942 0.875 0.977
DN-DETR CVPR’22 0.81 0.867 0.906 0.93 0.935 0.942 0.968 0.945 0.883 0.981
DN DEF. CVPR’22 0.815 0.867 0.898 0.933 0.944 0.964 0.990 0.950 0.892 0.98
FocalNet-DINO NeurIPS’22 0.823 0.893 0.929 0.956 0.964 0.979 0.992 0.950 0.893 0.985
Proposed 0.825 0.895 0.930 0.958 0.964 0.979 0.992 0.951 0.893 0.986

Abbreviations: AUC, area under the curve; R, recall; RCC, renal cell carcinoma.

Renal Cell Carcinoma Classification box with the highest confidence. A bounding box was con-
The RCC class (ccRCC vs. non-ccRCC) of a CT slice was sidered valid if its confidence was above a threshold of 0.5. If
determined according to the label of the valid bounding no valid bounding boxes were detected in a slice, that slice
was considered normal. The results of the classification
approach (sensitivity, specificity, F1 score, accuracy, and
AUC score) are included in ►Table 1. The precision recall
curve (PR curve) and the AUC curve are shown in ►Figs. 5
and 6, respectively. The FocalNet-DINO demonstrated an
accuracy of 0.950, an F1 score of 0.893, and an AUC score
of 0.985 for classification of RCC (ccRCC vs. non-ccRCC). The
addition of the spatial and class consistency modules led to
an improvement of 0.1% in the AUC score.

Discussion
In our study, we showed high accuracy of computationally
efficient 2D neural networks for detection and classification
of RCC on CT images by approaching it as an image-based
object detection task (best performance for FocalNet-DINO)
showing further improvement of the performance by incor-
Fig. 3 Free-response receiver operating characteristic (FROC) curve
comparing various state-of-the-art object detection architectures for poration of spatial and class consistency modules in our
renal cell carcinoma detection with the proposed model. proposed algorithm.

Fig. 4 Visualizations of the ground truth (GT) bounding boxes (blue) and predictions (green) for various tested algorithms for different patients.

Indian Journal of Radiology and Imaging Vol. 35 No. 3/2025 © 2024. Indian Radiological Association. All rights reserved.
400 Predicting Renal Cell Carcinoma Subtypes and Grading Using 2D Neural Networks Gupta et al.

works demonstrate an enhanced ability to handle data

variability, require less data for training and validation,
and offer faster inference times compared to their 3D
counterparts.18
Since the proposed 2D neural network analyses each slice
of the CT scan independently, there is a high risk of spurious
misclassification, false-positive box prediction, or deviated
box prediction. This error resembles the salt and pepper
noise of the images.20 To mitigate the above-mentioned
effects, we incorporated spatial consistency and class con-
sistency modules. We analyze the spatial overlap of the
predicted bounding boxes across the slices and replace the
bounding box of any slice with the median of the bounding
boxes in its N-5 neighborhood (5 previous and next frames).
This removes the false positives and adjusts the deviated
bounding boxes. Similarly, we incorporate the class consis-
tency module to replace the class of the detected bounding
Fig. 5 Precision recall (PR) curve comparing performance of various
state-of-the-art object detection architectures for renal cell carcino- box as a median of the classes predicted in the N-5 neigh-
ma classification with the proposed model. borhood. This reduces the risk of misclassification and
establishes consistency in the prediction throughout the
We used 2D neural networks for RCC diagnosis on CT patient’s CT scan. Incorporating these modules ensures the
images. While three-dimensional (3D) networks have the model is practically usable in the radiology workflow.
potential to offer a more comprehensive analysis by consid- Previous research has investigated the use of DL models
ering the volumetric context and spatial relationships within for image analysis of renal tumors, including tumor detec-
the entire tumor, these networks treat the entire CT scan of a tion, segmentation, and subclassification, with promising
patient as a single training example, resulting in requirement results.21–24 However, the majority of prior investigations
of a substantially large dataset for training.17,18 In scenarios have focused on distinguishing between benign and malig-
with limited data, as observed in our work, 3D convolutional nant renal tumors and cysts.25 Moreover, in these previous
neural networks (CNNs) are prone to overfitting, rendering studies, the DL methods for RCC classification required
them practically unusable.19 In contrast, 2D networks use manual tumor localization using tumor regions marked by
individual slices of the CT scan as training samples. This a radiologist.24,26,27 To the best of our knowledge, two
approach allows for training with a limited number of studies report the end-to-end use of DL for both RCC detec-
patient data. Additionally, due to the considerable diversity tion and its subtyping into histopathological variants.11,13
among CT slices, 2D CNNs exhibit a reduced likelihood of Both these studies used 3D CNN–based models and found
overfitting.19 Furthermore, 2D CNNs are more computation- high accuracy of these models for RCC classification, compa-
ally efficient, resulting in quicker processing times. This rable to the radiologists. Our study further contributes to the
efficiency is particularly pertinent in environments with literature by demonstrating similar high performance for
limited computational resources. Notably, 2D neural net- RCC detection and classification by an ingenious approach
using 2D neural networks and feature consistency modules.
Furthermore, this study adds to the expanding body of
evidence supporting the use of DNN in medical imaging,
notably for accurately and efficiently classifying complicated
malignancies such as RCC. Such differentiation of the more
aggressive ccRCC from other non-ccRCC tumor subtypes has
the potential to improve prognosis and enable personalized
treatments with targeted therapies like tyrosine kinase and
vascular endothelial growth factor inhibitors. Given the
limitations of the current diagnostic methods in reliably
achieving this differentiation, the use of DL models offers a
promising noninvasive alternative for accurate RCC subtyp-
ing and grading, potentially improving patient outcomes and
treatment precision.
Our study had certain limitations. Our study relied on
postoperative pathology reports, which limited our analysis
to nonmetastatic cases, potentially affecting the generaliz-
Fig. 6 Area under the curve (AUC) comparing classification perfor-
ability of our findings. Metastatic RCCs, especially non-ccRCC
mance of various state-of-the-art object detection architectures with subtypes, can exhibit heterogeneity, necrosis, and other
the proposed model. features distinct from nonmetastatic cases, potentially

Indian Journal of Radiology and Imaging Vol. 35 No. 3/2025 © 2024. Indian Radiological Association. All rights reserved.
 Predicting Renal Cell Carcinoma Subtypes and Grading Using 2D Neural Networks Gupta et al. 401

affecting the DL features extracted. In the cases where the 9 Young JR, Margolis D, Sauk S, Pantuck AJ, Sayre J, Raman SS. Clear
model failed to detect or classify RCC accurately, contributing cell renal cell carcinoma: discrimination from other renal cell
factors likely included low tumor-to-tissue contrast, hetero- carcinoma subtypes and oncocytoma at multiphasic multidetec-
tor CT. Radiology 2013;267(02):444–453
geneous or necrotic tumor regions, and reduced spatial
10 Xu L, Yang C, Zhang F, et al. Deep learning using CT images to grade
coherence in certain slices. A comprehensive failure analysis clear cell renal cell carcinoma: development and validation of a
could help identify specific patterns in misclassifications and prediction model. Cancers (Basel) 2022;14(11):2574
inform future model improvements. Due to the small indi- 11 Yao N, Hu H, Chen K, et al. A robust deep learning method with
vidual sample sizes for chromophobe and papillary RCC uncertainty estimation for the pathological classification of renal
cell carcinoma based on CT images. J Imaging Inform Med 2025;38
subgroups, a meaningful analysis of subgroup differences
(03):1323–1333
within the non-ccRCC class could not be performed. The
12 Wang Z, Zhang X, Wang X, et al. Deep learning techniques for
long-term impact of integrating our method into clinical imaging diagnosis of renal cell carcinoma: current and emerging
practice remains to be seen. Future research should look into trends. Front Oncol 2023;13:1152622
the practical usefulness of this method in clinical settings. 13 Uhm KH, Jung SW, Choi MH, et al. Deep learning for end-to-end
Integrating our method with other diagnostic modalities, kidney cancer diagnosis on multi-phase abdominal computed
tomography. NPJ Precis Oncol 2021;5(01):54
such as MRI or ultrasound, could improve the accuracy and
14 Amador S, Beuschlein F, Chauhan V, et al. Deep learning
usefulness of RCC classification. approaches applied to image classification of renal tumors: a
systematic review. Arch Comput Methods Eng 2024;31(02):
615–622
Conclusion 15 Dai C, Xiong Y, Zhu P, et al. Deep learning assessment of small
In conclusion, in this study, we leverage 2D neural networks renal masses at contrast-enhanced multiphase CT. Radiology
2024;311(02):e232178
and spatial and class consistency modules to offer a novel,
16 Mahootiha M, Qadir HA, Bergsland J, Balasingham I. Multimodal
computationally simpler and accurate DL approach to RCC deep learning for personalized renal cell carcinoma prognosis:
detection and classification on cross-sectional imaging. integrating CT imaging and clinical data. Comput Methods Pro-
grams Biomed 2024;244:107978
Funding 17 Yang YQ, Guo YX, Xiong JY, et al. Swin3d: a pretrained transformer
backbone for 3d indoor scene understanding. arXiv preprint
None.
arXiv:2304.06906. 2023
18 Zhang Y, Shi L, Wu Y, Cheng K, Cheng J, Lu H. Gesture recognition
Conflict of Interest based on deep deformable 3D convolutional neural networks.
None declared. Pattern Recognit 2020;107:107416
19 Ying X. An overview of overfitting and its solutions. J Phys Conf Ser
2019;1168(02):022022
References 20 Zheng Q, Delingette H, Duchateau N, Ayache N. 3-D consistent and
1 Moch H, Cubilla AL, Humphrey PA, Reuter VE, Ulbright TM. The robust segmentation of cardiac images by deep learning with
2016 WHO classification of tumours of the urinary system and spatial propagation. IEEE Trans Med Imaging 2018;37(09):
male genital organs-part A: renal, penile, and testicular tumours. 2137–2148
Eur Urol 2016;70(01):93–105 21 Tanaka T, Huang Y, Marukawa Y, et al. Differentiation of small
2 Capitanio U, Cloutier V, Zini L, et al. A critical assessment of the ( 4 cm) renal masses on multiphase contrast-enhanced CT by
prognostic value of clear cell, papillary and chromophobe histo- deep learning. AJR Am J Roentgenol 2020;214(03):605–612
logical subtypes in renal cell carcinoma: a population-based 22 Oberai A, Varghese B, Cen S, et al. Deep learning based classifica-
study. BJU Int 2009;103(11):1496–1500 tion of solid lipid-poor contrast enhancing renal masses using
3 Patard JJ, Leray E, Rioux-Leclercq N, et al. Prognostic value of contrast enhanced CT. Br J Radiol 2020;93(1111):20200002
histologic subtypes in renal cell carcinoma: a multicenter experi- 23 Zabihollahy F, Schieda N, Krishna S, Ukwatta E. Automated
ence. J Clin Oncol 2005;23(12):2763–2771 classification of solid renal masses on contrast-enhanced com-
4 Lee CH, Motzer R. Combination VEGFR/immune checkpoint in- puted tomography images using convolutional neural network
hibitor therapy: a promising new treatment for renal cell carci- with decision fusion. Eur Radiol 2020;30(09):5183–5190
noma. Br J Cancer 2018;119(08):911–912 24 Han S, Hwang SI, Lee HJ. The classification of renal cancer in 3-
5 Atkins MB, Tannir NM. Current and emerging therapies for first- phase CT images using a deep learning method. J Digit Imaging
line treatment of metastatic clear cell renal cell carcinoma. Cancer 2019;32(04):638–643
Treat Rev 2018;70:127–137 25 Amador S, Beuschlein F, Chauhan V, et al. Deep learning
6 Abel EJ, Carrasco A, Culp SH, et al. Limitations of preoperative approaches applied to image classification of renal tumors. a
biopsy in patients with metastatic renal cell carcinoma: compari- systematic review. Arch Computat Methods Eng 2024;31:
son to surgical pathology in 405 cases. BJU Int 2012;110(11): 615–622
1742–1746 26 Zheng Y, Wang S, Chen Y, Du HQ. Deep learning with a convolu-
7 Abel EJ, Culp SH, Matin SF, et al. Percutaneous biopsy of primary tional neural network model to differentiate renal parenchymal
tumor in metastatic renal cell carcinoma to predict high risk tumors: a preliminary study. Abdom Radiol (NY) 2021;46(07):
pathological features: comparison with nephrectomy assess- 3260–3268
ment. J Urol 2010;184(05):1877–1881 27 Zuo T, Zheng Y, He L, et al. Automated classification of papillary
8 Karlo CA, Di Paolo PL, Chaim J, et al. Radiogenomics of clear cell renal cell carcinoma and chromophobe renal cell carcinoma based
renal cell carcinoma: associations between CT imaging features on a small computed tomography imaging dataset using deep
and mutations. Radiology 2014;270(02):464–471 learning. Front Oncol 2021;11:746750

Indian Journal of Radiology and Imaging Vol. 35 No. 3/2025 © 2024. Indian Radiological Association. All rights reserved.