Utilizing Radiomic Features for Automated MRI Keypoint Detection:

Enhancing Graph Applications

Sahar Almahfouz Nasser

, Shashwat Pathak

, Keshav Singhal

, Mohit Meena

Nihar Gupte

, Ananya Chinmaya

, Prateek Garg

and Amit Sethi

Department of Electrical Engineering, Indian Institute of Technology Bombay, Mumbai, India

Keywords:

Image Matching, Image Registration, Kepoint Detection, Radiomic Features, Brain MRI, GNN.

Abstract:

Graph neural networks (GNNs) present a promising alternative to CNNs and transformers for certain image

processing applications due to their parameter-efﬁciency in modeling spatial relationships. Currently, an active

area of research is to convert image data into graph data as input for GNN-based models. A natural choice for

graph vertices, for instance, are keypoints in images. SuperRetina is a promising semi-supervised technique

for detecting keypoints in retinal images. However, its limitations lie in the dependency on a small initial

set of ground truth keypoints, which is progressively expanded to detect more keypoints. We encountered

difﬁculties in detecting a consistent set of initial keypoints in brain images using traditional keypoint detection

techniques, such as SIFT and LoFTR. Therefore, we propose a new approach for detecting the initial keypoints

for SuperRetina, which is based on radiomic features. We demonstrate the anatomical signiﬁcance of the

detected keypoints by showcasing their efﬁcacy in improving image registration guided by these keypoints.

We also employed these keypoints as ground truth for a modiﬁed keypoint detection method known as LK-

SuperRetina for improved image matching in terms of both the number of matches and their conﬁdence scores.

1 INTRODUCTION

Graph neural networks (GNNs) have shown promis-

ing results for reducing the computational require-

ments for certain image processing tasks as shown

in (He et al., 2023). However, converting images into

graphs is an active area of research. One approach is

to break down an image into patches and treat each

patch as a node in a graph while another is to segment

an image into super-pixels and use each super-pixel

as a graph vertex (He et al., 2023). Our method relies

on detecting important keypoints in the images along

with their features and making graphs.

Detecting important keypoints in certain types of

images, such as magnetic resonance images of the

brain is not straightforward due to the lack of well-

https://orcid.org/0000-0002-5063-9211

https://orcid.org/0000-0002-4508-8531

https://orcid.org/0009-0003-2290-4924

https://orcid.org/0009-0008-1641-2727

https://orcid.org/0009-0004-5972-2088

https://orcid.org/0009-0008-5844-5498

https://orcid.org/0009-0005-5348-8905

https://orcid.org/0000-0002-8634-1804

deﬁned landmarks. In various registration competi-

tions, such as the BraTSReg challenge (Baheti et al.,

2021), experts have had to mark speciﬁc landmark

points on 3D brain images to help the participants

align images and evaluate how well their registration

algorithms worked. But this marking is neither easy

nor fast. Also, only a few points – between 6 and 50

per set of images – get marked, which can be inade-

quate for good registration.

Traditional keypoint detection algorithms such as

SIFT (Lindeberg, 2012) fall behind deep learning-

based keypoint detection algorithms which have dif-

ferent types include: supervised, unsupervised, and

semi-supervised methods. Some examples are Un-

superPoint (Christiansen et al., 2019), SuperPoint

(DeTone et al., 2018), GLAMpoints (Truong et al.,

2019), and SuperRetina (Liu et al., 2022).

After trying and failing to ﬁnd reliable keypoints

in brain images using methods such as SIFT (Linde-

berg, 2012) and LoFTR (Sun et al., 2021), we pro-

posed a new algorithm. Our proposed method ﬁnds

keypoints and extract their descriptors using radiomic

features in brain MR images. We assessed the util-

ity of these keypoints in brain images by using them

for image registration. Then we used a dataset that

Nasser, S., Pathak, S., Singhal, K., Meena, M., Gupte, N., Chinmaya, A., Garg, P. and Sethi, A.

Utilizing Radiomic Features for Automated MRI Keypoint Detection: Enhancing Graph Applications.

DOI: 10.5220/0012568800003657

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 17th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2024) - Volume 1, pages 319-325

ISBN: 978-989-758-688-0; ISSN: 2184-4305

319

we prepared, which consists of MRI images from

OASIS dataset (Marcus et al., 2007) and the key-

points we detected, to train the LK-SuperRetina al-

gorithm (Almahfouz Nasser et al., 2023) to detect

new keypoints. We showed that using a GNN, such

as SuperGlue (Sarlin et al., 2020), for image match-

ing improves the features of detected keypoints. This

method of detecting keypoints opens doors for using

graph-based neural networks for different tasks, in-

cluding brain classiﬁcation, segmentation, and regis-

tration.

Our contribution can be outlined as follows:

• Introduction of an innovative, fully automated ap-

proach for keypoint detection in brain MRI im-

ages.

• Development of a two-stage method: the ﬁrst

stage involves detecting initial keypoints by uti-

lizing segmentations of regions of interest in the

brain, employing radiomic features as keypoint

descriptors, and establishing keypoints at the cen-

ters of mass of the segmentations. The sec-

ond stage employs our previously developed algo-

rithm, LK-SuperRetina, gradually increasing the

number of detected keypoints, commencing from

the initial set.

• Enhancement of various tasks, including im-

age registration and image matching, achieved

through the incorporation of detected keypoints.

2 RELATED WORK

Traditional approaches to keypoint detection have

long been prominent in the ﬁeld of computer vision,

with a focus on identifying keypoints in images that

remain consistent across scaling, rotation, and light-

ing variations. These techniques involve character-

izing the local image patch around these keypoints

using a set of features, allowing for the matching of

keypoints between different images and object recog-

nition. However, these methods come with their limi-

tations, including high computational complexity, re-

duced accuracy in the presence of extreme lighting

and viewpoint changes, and challenges in handling

occlusions and cluttered backgrounds.

In recent years, deep learning-based keypoint de-

tection algorithms have emerged as a promising al-

ternative. These algorithms have the ability to au-

tonomously learn robust and discriminative features

directly from data, making them better suited to han-

dle complex and diverse image variations. As a result,

they have found application in various domains, in-

cluding object detection, semantic segmentation, and

image retrieval.

Within the realm of deep learning, keypoint de-

tection algorithms come in various forms, including

supervised, semi-supervised, self-supervised, and un-

supervised techniques. Supervised techniques require

annotated data, where keypoints are manually labeled

in training images, proving advantageous in scenar-

ios with a substantial volume of labeled data, such

as facial recognition or object detection. Conversely,

unsupervised techniques operate independently of la-

beled data, with the network learning to identify key-

points by maximizing speciﬁc objectives, like infor-

mation preservation during feature extraction. These

methods are particularly valuable in contexts where

obtaining labeled data is challenging or expensive,

such as in medical imaging or remote sensing.

Prominent deep learning-based keypoint detection

methods include UnsuperPoint, SuperPoint, GLAM-

points, and SuperRetina. UnsuperPoint (Christiansen

et al., 2019) introduces an innovative unsupervised

training approach, utilizing a combination of differ-

entiable soft nearest neighbor loss and unsupervised

clustering loss. SuperPoint (DeTone et al., 2018),

on the other hand, represents a self-supervised deep

learning-based algorithm for keypoint detection and

description, relying on a novel loss function for train-

ing on unlabeled images, which makes it adaptable

and scalable across various applications.

The primary distinction between UnsuperPoint

and SuperPoint lies in their training methodologies.

While SuperPoint follows a self-supervised approach,

UnsuperPoint takes an unsupervised path. Addition-

ally, UnsuperPoint achieves state-of-the-art perfor-

mance across various benchmarks, even outperform-

ing SuperPoint in challenging scenarios characterized

by signiﬁcant viewpoint alterations and illumination

shifts. GLAMpoints (Truong et al., 2019) emerges as

a semi-supervised deep learning-based algorithm for

interest point detection and description, employing a

unique greedy training strategy for end-to-end learn-

ing of keypoint detection and description. This strat-

egy involves learning to select the most precise key-

points and their descriptors, resulting in higher accu-

racy and efﬁciency, especially in demanding scenarios

involving signiﬁcant viewpoint changes, scaling, ro-

tation, and benchmarks like HPatches (Balntas et al.,

2017). Furthermore, GLAMpoints is well-suited for

accommodating multiple object instances within the

same image, making it suitable for multi-object track-

ing and matching. SuperRetina (Liu et al., 2022) rep-

resents a semi-supervised approach for keypoint de-

tection and description in retinal images, leveraging

both labeled and unlabeled data to enhance the per-

formance of the keypoint detector and descriptor.

BIOIMAGING 2024 - 11th International Conference on Bioimaging

320

In the following section, we introduce our pro-

posed approach for detecting keypoints in brain MRI

images. As we will demonstrate, this task proves to be

highly challenging, and traditional keypoint detection

methods like SIFT, as well as well-known deep learn-

ing techniques such as LoFTR, struggle to identify ro-

bust and repeatable keypoints. Our solution involves

leveraging radiomic features for initial keypoint de-

tection in the brain. Subsequently, we employ LK-

SuperRetina to increase the number of identiﬁed key-

points in the images. Finally, we showcase the effec-

tiveness of these keypoints in applications like image

matching and registration, highlighting their positive

impact on the overall performance of these tasks.

3 PROPOSED METHOD

We now introduce our approach for keypoint detec-

tion using radiomic features. Given an image along-

side its segmentation labels, our method identiﬁes ra-

diomic keypoints as the centers of radiomic segmen-

tation labels within the image. These radiomic fea-

tures encompass a range of intensity characteristics

(such as mean intensity, standard deviation, skew-

ness and kurtosis), shape characteristics (such as vol-

ume, surface area, and compactness), texture char-

acteristics (such as gray level co-occurrence matrix

(GLCM), gray level run length matrix (GLRLM), and

gray level run length matrix)), wavelet-Based charac-

teristics, spatial and statistical characteristics (such as

centroid position, eccentricity, and entropy), and frac-

tal characteristics. These radiomic features are spe-

ciﬁc to the regions deﬁned by the segmentation map.

To compute these features, we utilized the Pyradiomic

library (Van Griethuysen et al., 2017).

Radiomic keypoints are closely tied to segmenta-

tion regions predicted by neural networks. We trained

Swin UNetR (Hatamizadeh et al., 2021) to predict the

segmentation maps of brain images and used them as

masks to extract the keypoints. Each segmentation

mask yields a keypoint location, accompanied by 53

descriptive radiomic features, see ﬁgures 1 and 2 for

more details. Radiomic keypoints exhibit repeatabil-

ity across various brain samples even in the presence

of varying intensity and non-rigid deformations which

make our proposed method a more robust alternative

to methods, such as SIFT (Lindeberg, 2012). These

detected keypoints can serve as a initial keypoints for

detecting additional keypoints through deep learning

techniques as we will show in our results, in which

we trained LK-SuperRetina (Almahfouz Nasser et al.,

2023) to automatically detect these ground truth key-

points and extra keypoints. This underscores the sig-

niﬁcance of these keypoints as essential landmarks

within brain images.

4 EXPERIMENTS AND RESULTS

This section begins with an introduction to the dataset

used in all our experiments. Following that, we ex-

plore various approaches applied for keypoint detec-

tion, explaining how the detected keypoints were uti-

lized to improve registration and identify additional

keypoints. Finally, we will conclude with our results

on GNN-based image matching.

4.1 OASIS Dataset

The OASIS dataset, as described in (Marcus et al.,

2007), contains MRI data obtained from 414 subjects.

This dataset has been divided into three separate sub-

sets for training, validation, and testing, following a

ratio of 314:50:50, respectively. Each subject in the

dataset has T1-weighted scan, as well as segmentation

masks of various regions of the brain. This dataset

incorporates three distinct types of brain segmenta-

tion: a four-label mask, a thirty-ﬁve label mask, and

a twenty-four label mask.The 3D T1-weighted scans

and their corresponding masks are of resolution (160,

192, 224). Additionally, the 2D T1-weighted scans

and their corresponding masks have a resolution of

(160, 192).

4.2 SIFT

SIFT (Lindeberg, 2012) was initially chosen for ex-

perimentation due to its reputation in keypoint detec-

tion. Utilizing gradients, SIFT excels in identifying

scale and rotation invariant keypoints. The method

calculates orientation gradients across scales, creat-

ing a distinctive 128-dimensional vector for each key-

point.

Figure 3 shows signiﬁcant issues we encountered

with keypoint detection using SIFT. Notably, there

was inconsistency in keypoint locations across differ-

ent MRI slices within the brain, reﬂecting a lack of

repeatability. Our experiments highlighted SIFT’s in-

effectiveness under conditions involving large defor-

mations.

4.3 LoFTR

Due to SIFT’s inability to establish correspondences,

we opted to explore a deep learning method namely

LoFTR (Sun et al., 2021), which stands for detector-

free local feature matching transformer.

Utilizing Radiomic Features for Automated MRI Keypoint Detection: Enhancing Graph Applications

321

Figure 1: The image matching pipeline, encompassing keypoint detection using neural networks, graph formation from

detected keypoints, graph neural network (GNN) processing to enhance keypoint features, and a dedicated head for keypoint

matching.

Segmentation

Network

Radiomic

Block

Radiomic

Block

Radiomic

Block

Radiomic

Block

x1,y1,z1

Features

x2,y2,z2

Features

x3,y3,z3

Features

xn,yn,zn

Features

Input Image

Segmentation

Output

Figure 2: Radiomic features-based Keypoint detection

method.

In the LoFTR approach, a convolutional neural

network (CNN) with an encoder-decoder architecture

extracts both low-resolution and high-level features.

The LoFTR module, incorporating self-attention and

cross-attention blocks, transforms these features, and

a differential matching layer offers two methods: op-

timal transport and dual softmax (Bridle, 1989) .

Our implementation used a pretrained model from

the Aachen Day Night Dataset (Sattler et al., 2018).

LoFTR succeeded in generating keypoint matches for

similar intensity proﬁles but faced challenges in other

cases similar to SIFT.These ﬁndings emphasize the

need for keypoints exhibiting consistency, accurate

matching, and intensity proﬁle invariance across the

dataset.

Figure 3: SIFT (Lindeberg, 2012) and LoFTR (Sun et al.,

2021) performances in keypoint detection. The left pair of

images depicts results for SIFT, while the right one illus-

trates the performance of LoFTR. It is evident that LoFTR

outperforms SIFT in keypoint detection and matching. Nev-

ertheless, LoFTR encounters challenges when the source

and target images exhibit varying intensity distributions.

4.4 Radiomic Features-Based Keypoint

Detection

Having faced challenges with keypoint detection us-

ing SIFT and LoFTR, which proved sensitive to inten-

sity variations, we shifted to our proposed radiomic

features-based method. In this section, we prove the

importance of the detected keypoints by showing our

results on registering brain images among different

subjects. we also showcase the applications of the

dataset containing the original OASIS scans and the

corresponding detected radiomic keypoints on auto-

matic keypoint detection. And ﬁnally train a GNN-

based image matcher on the same dataset.

4.4.1 Image Registration

To assess the signiﬁcance of radiomics keypoints as

landmarks, we integrated them into the loss function

of a registration network (TransMorph (Chen et al.,

2022)). Our ﬁndings revealed that incorporating the

keypoints’ loss led to a notable 3% enhancement in

the registration performance.

Vision transformers, excelling in capturing long-

range spatial relationships, prove effective in medical

image tasks due to their large receptive ﬁelds. Trans-

Morph (Chen et al., 2022), a hybrid Transformer-

BIOIMAGING 2024 - 11th International Conference on Bioimaging

322

ConvNet model, utilizes these advantages for volu-

metric medical image registration. The encoder di-

vides input volumes into 3D patches, projecting them

to feature representations through linear layers. Se-

quential patch merging and Swin Transformer blocks

follow. The decoder, with upsampling and convolu-

tional layers, connects to the encoder stages via skip

connections, producing the deformation ﬁeld. We

contributed by designing a customized loss function

for keypoints, utilizing Gaussian-blurred keypoints to

create a ground truth heatmap. Combining Dice and

inverted Dice losses addressed imbalanced masks, re-

sulting in a 3% Dice score improvement over the OA-

SIS test dataset. TransMorph achieved a dice score

of 0.89 with keypoint loss, compared to 0.86 with-

out, underscoring keypoints’ role in enhancing reg-

istration performance. A potential avenue for future

research involves developing a loss function that con-

siders both the feature descriptors of keypoints and

the disparity between the locations of registered key-

points and their counterparts in the target image.

4.4.2 Automated Keypoint Detection

SuperRetina, introduced in (Liu et al., 2022) is an ap-

daptive version of SuperPoint model (DeTone et al.,

2018) for identifying important keypoints in retinal

images. Utilizing a semi-supervised learning frame-

work, SuperRetina maximizes the utility of limited

labeled retinal image data by combining both super-

vised and unsupervised techniques. Yet, its utilization

requires an initial set of ground truth keypoints to ini-

tiate the process, subsequently increasing the detected

keypoints iteratively. In our approach, we use our ra-

diomic keypoints as the initial sets for OASIS images.

LK-SuperRetina (Almahfouz Nasser et al., 2023)

which is a modiﬁed version of SuperRetina consists

of an encoder for downsampling, along with two de-

coders—one for keypoint detection and another for

descriptor generation. Keypoint detection utilizes a

mix of labeled and unlabeled data, while descriptor

training employs self-supervised learning.

Following the U-Net (Ronneberger et al., 2015)

design, LK-SuperRetina’s shallow encoder begins

with a single convolutional layer, followed by three

blocks containing two convolutional layers, a 2 × 2

max-pooling layer, and ReLU activation. The key-

point decoder has three blocks with two convolutional

layers, ReLU activation, and concatenation block.

The detection map (P) is generated through a convo-

lutional block with three convolutional layers and a

sigmoid activation.

The loss function combines the detector and the

descriptor losses. The detection loss consists of the

classiﬁcation loss and the geometric loss as shown in

Equation 1.

det

= l

cl f

+ l

geo

(1)

The classiﬁcation loss component (l

cl f

) is deﬁned

in Equation 2, where

Y represents the smoothed ver-

sion of the binary ground truth labels Y of the key-

points after blurring them with a 2D Gaussian.

cl f

(I;Y ) = 1 −

∑

i, j

(P ◦

Y )

i, j

∑

i, j

(P ◦ P)

i, j

∑

i, j

(

Y ◦

Y )

i, j

(2)

Where ◦ denotes element-wise multiplication.

The detector generates a heatmap as its output.

Coordinates where the intensity value exceeds a spec-

iﬁed threshold (subject to non-maximum suppression)

are regarded as the keypoint coordinates. When feed-

ing both the image I and its augmented version I

′

the network, two tensors for the descriptors D and

′

are obtained. For each keypoint (i, j) in the non-

maximum suppressed keypoint set

P, two distances

are computed: Φ

rand

i, j

between the descriptors of (i, j)

in the set

P and a random point from the registered

heatmap H(

P), and Φ

hard

i, j

representing the minimal

distance, as depicted in Equation 3.

des

(I, H) =

∑

(i, j)∈

max



0, m + Φ

i, j

−



rand

i, j

+ Φ

hard

i, j





(3)

For more in-depth information on the loss func-

tion, please refer to the SuperRetina paper (Liu et al.,

2022). Figure 4 shows the results obtained from

LK-SuperRetina. As demonstrated, the number of

additionally identiﬁed keypoints meets expectations,

showcasing the network’s proﬁciency in capturing

good new keypoints. The network successfully de-

tects both the ground truth keypoints and extra key-

points during the testing phase.

Figure 4 presents two instances demonstrating the

resilience of the keypoint detection model against de-

formations. The images were deformed randomly by

an afﬁne deformation. We passed the original image

(target) and the deformed image (reference) seper-

ately to LK-SuperRetina. The model successfully

identiﬁed corresponding keypoints in both images, as

indicated by the number of good matches. We ad-

justed the thresholds of LK-SuperRetina to detect a

smaller set of keypoints for the clarity of the visual-

ization, but in practice, the model can detect over 300

good keypoints.

4.4.3 Image Matching

Following the detection of keypoints within the brain

images, we proceed to construct graphs to be used as

Utilizing Radiomic Features for Automated MRI Keypoint Detection: Enhancing Graph Applications

323

Table 1: A comparison between brute force matcher and SuperGlue. SuperGlue outperforms the brute force matcher in terms

of both evaluation metrics: the average number of good matches and the average conﬁdence score across the entire test dataset.

Method Avg. No. Good Matches Conﬁdence Score

BF (Lowe, 2004) 7 0.449 ± 0.007

SuperGlue 20 0.988 ± 0.010

Figure 4: Two examples show the robustness of the key-

point detection model. In both rows, the sequence of im-

ages, from left to right, includes: the reference image, the

target image, and the registration output. The good matches

contribute to aligning the images effectively.

inputs of the GNN, for accomplishing speciﬁc tasks

such as matching in our study. Within this paper,

we demonstrate our success in training a GNN-based

matcher (SuperGlue (Sarlin et al., 2020)) using the

graphs formed from the detected keypoints.

SuperGlue designed for matching two sets of lo-

cal features by identifying correspondences and ﬁl-

tering non-matchable points. Using attention-based

graph neural networks, it integrates context aggrega-

tion, matching, and ﬁltering within a uniﬁed architec-

ture. SuperGlue employs self-attention to enhance the

receptive ﬁeld of local descriptors and cross-attention

for cross-image communication. The network han-

dles partial assignments and occluded points by solv-

ing an optimal transport problem. With superior per-

formance over other learned approaches, SuperGlue

achieves state-of-the-art results in pose estimation for

challenging real-world indoor and outdoor environ-

ments.

Figure 5 and Table 1 show a performance compar-

ison between the brute force matcher and SuperGlue

across the test dataset. SuperGlue enhances the fea-

tures of the detected keypoints which improves the

matching performance. The data presented in the ta-

ble suggests that exploring the application of GNNs

for keypoint matching is a promising endeavor. To

pursue this, it is essential to identify keypoints in im-

ages and format the data to serve as appropriate input

for GNNs. Having introduced a technique for detect-

ing signiﬁcant keypoints in brain MRI images, these

keypoints can now function as nodes in graphs, pro-

viding data for GNNs. This development paves the

way for a multitude of graph-based applications in the

analysis of brain MRI images.

CV2 BF Matcher - Correct Matches = 7, Incorrect = 4, Failed = 9

CV2 BF Matcher - Correct Matches = 4, Incorrect = 6, Failed = 10 GNN Matcher - Good Matches = 20, Failed = 0, Avg. Confidence = 0.9526

GNN Matcher - Good Matches = 20, Failed = 0, Avg. Confidence = 0.9932

—— 0.90

—— 0.80

—— 0.70

—— 0.55

—— 0.40

—— 0.25

—— 0.10

Avg. Confidence Score Colour Spectrum

Figure 5: A comparison of the SuperGlue and brute force

matcher performance in matching detected keypoints on

brain images.

5 CONCLUSION

To sum up, our radiomic keypoint detection algorithm

provides a solution for automated keypoint detection

in MRI scans, overcoming challenges encountered by

traditional and other deep learning methods. The lim-

ited set of radiomic keypoints facilitates training Su-

perRetina for increased keypoint detection. These

keypoints are consistent and deformation resilient.

Our approach paves the way for the application of

GNN-based models on brain images, offering a faster

and more parameter-efﬁcient alternative compared to

CNNs and transformers. Moreover, the detection of

keypoints contributes to various tasks, including reg-

istration as justiﬁed in this work.

The limitation of this study is the necessity for

segmenting regions of interest to acquire radiomic

features, which are subsequently utilized as the initial

keypoints for the keypoint detection algorithm, LK-

SuperRetina.

REFERENCES

Almahfouz Nasser, S., Gupte, N., and Sethi, A. (2023). Re-

verse knowledge distillation: Training a large model

using a small one for retinal image matching on lim-

ited data. arXiv e-prints, pages arXiv–2307.

Baheti, B., Waldmannstetter, D., Chakrabarty, S., Akbari,

H., Bilello, M., Wiestler, B., Schwarting, J., Cal-

abrese, E., Rudie, J., Abidi, S., et al. (2021). The brain

tumor sequence registration challenge: establishing

correspondence between pre-operative and follow-up

mri scans of diffuse glioma patients. arXiv preprint

arXiv:2112.06979.

Balntas, V., Lenc, K., Vedaldi, A., and Mikolajczyk, K.

BIOIMAGING 2024 - 11th International Conference on Bioimaging

324

(2017). Hpatches: A benchmark and evaluation of

handcrafted and learned local descriptors. In Proceed-

ings of the IEEE conference on computer vision and

pattern recognition, pages 5173–5182.

Bridle, J. (1989). Training stochastic model recognition al-

gorithms as networks can lead to maximum mutual in-

formation estimation of parameters. Advances in neu-

ral information processing systems, 2.

Chen, J., Frey, E. C., He, Y., Segars, W. P., Li, Y., and Du,

Y. (2022). Transmorph: Transformer for unsupervised

medical image registration. Medical image analysis,

82:102615.

Christiansen, P. H., Kragh, M. F., Brodskiy, Y., and

Karstoft, H. (2019). Unsuperpoint: End-to-end unsu-

pervised interest point detector and descriptor. arXiv

preprint arXiv:1907.04011.

DeTone, D., Malisiewicz, T., and Rabinovich, A. (2018).

Superpoint: Self-supervised interest point detection

and description. In Proceedings of the IEEE con-

ference on computer vision and pattern recognition

workshops, pages 224–236.

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H. R.,

and Xu, D. (2021). Swin unetr: Swin transformers

for semantic segmentation of brain tumors in mri im-

ages. In International MICCAI Brainlesion Workshop,

pages 272–284. Springer.

He, X., Hooi, B., Laurent, T., Perold, A., LeCun, Y., and

Bresson, X. (2023). A generalization of vit/mlp-mixer

to graphs. In International Conference on Machine

Learning, pages 12724–12745. PMLR.

Lindeberg, T. (2012). Scale invariant feature transform.

Liu, J., Li, X., Wei, Q., Xu, J., and Ding, D. (2022). Semi-

supervised keypoint detector and descriptor for retinal

image matching. In Computer Vision–ECCV 2022:

17th European Conference, Tel Aviv, Israel, October

23–27, 2022, Proceedings, Part XXI, pages 593–609.

Springer.

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60:91–110.

Marcus, D. S., Wang, T. H., Parker, J., Csernansky, J. G.,

Morris, J. C., and Buckner, R. L. (2007). Open ac-

cess series of imaging studies (oasis): cross-sectional

mri data in young, middle aged, nondemented, and

demented older adults. Journal of cognitive neuro-

science, 19(9):1498–1507.

Ronneberger, O., Fischer, P., and Brox, T. (2015). U-

net: Convolutional networks for biomedical image

segmentation. In Medical Image Computing and

Computer-Assisted Intervention–MICCAI 2015: 18th

International Conference, Munich, Germany, October

5-9, 2015, Proceedings, Part III 18, pages 234–241.

Springer.

Sarlin, P.-E., DeTone, D., Malisiewicz, T., and Rabinovich,

A. (2020). Superglue: Learning feature matching

with graph neural networks. In Proceedings of the

IEEE/CVF conference on computer vision and pattern

recognition, pages 4938–4947.

Sattler, T., Maddern, W., Toft, C., Torii, A., Hammarstrand,

L., Stenborg, E., Safari, D., Okutomi, M., Pollefeys,

M., Sivic, J., et al. (2018). Benchmarking 6dof out-

door visual localization in changing conditions. In

Proceedings of the IEEE conference on computer vi-

sion and pattern recognition, pages 8601–8610.

Sun, J., Shen, Z., Wang, Y., Bao, H., and Zhou, X. (2021).

Loftr: Detector-free local feature matching with trans-

formers. In Proceedings of the IEEE/CVF conference

on computer vision and pattern recognition, pages

8922–8931.

Truong, P., Apostolopoulos, S., Mosinska, A., Stucky, S.,

Ciller, C., and Zanet, S. D. (2019). Glampoints:

Greedily learned accurate match points. In Proceed-

ings of the IEEE/CVF International Conference on

Computer Vision, pages 10732–10741.

Van Griethuysen, J. J., Fedorov, A., Parmar, C., Hosny, A.,

Aucoin, N., Narayan, V., Beets-Tan, R. G., Fillion-

Robin, J.-C., Pieper, S., and Aerts, H. J. (2017).

Computational radiomics system to decode the radio-

graphic phenotype. Cancer research, 77(21):e104–

e107.

Utilizing Radiomic Features for Automated MRI Keypoint Detection: Enhancing Graph Applications

325