Categories
News

Artificial intelligence-assisted magnetic resonance imaging technology in the differential diagnosis and prognosis prediction of endometrial cancer


Mannequin development

A picture processing mannequin primarily based on deep studying residual community was constructed. Based mostly on ResNet-10129, consideration mechanism was launched into the mannequin to enhance the focus of the mannequin on the area of curiosity. The eye module contains spatial consideration and channel consideration, that are mixed to enhance the characteristic illustration capacity of the mannequin. The general operating framework of the mannequin is illustrated in Fig. 1.

Fig. 1
figure 1

General operational framework of the improved mannequin.

To make the mannequin pay extra consideration to the essential channel options in the characteristic graph tensor, a channel consideration module is added to every node of the ResNet-101 structure to acquire a brand new residual module (Fig. 2). The operation steps of this channel consideration residual module are as follows: first, the picture options are processed by world common pooling, and then the generated characteristic photos are compressed. Then, two totally linked layers are used to mannequin the relationship between the channels, in order that the weight of the output options is in keeping with that of the enter options. Sigmoid activation perform was used to normalize the weight, and the obtained weight was used to symbolize the consideration degree of every characteristic channel. The characteristic weight worth of the channel was normalized.

$$:textual content{Z}={upphi:}left(textual content{M}proper(textual content{A}textual content{v}textual content{e}textual content{r}textual content{a}textual content{g}textual content{e}:textual content{p}textual content{o}textual content{o}textual content{l}textual content{i}textual content{n}textual content{g}left(textual content{L}proper)left)proper)+textual content{M}left(textual content{M}textual content{a}textual content{x}textual content{i}textual content{m}textual content{u}textual content{m}:textual content{p}textual content{o}textual content{o}textual content{l}proper(textual content{L}left)proper))$$

(1)

$$:textual content{L}textual content{*}=textual content{Z}diamond:textual content{L}$$

(2)

Z represents consideration weight, L represents enter characteristic, (:Ltext{*}) represents weighted characteristic, (:{upphi:})() represents ReLU activation perform, M means multi-layer perceptron, (:textual content{A}textual content{v}textual content{e}textual content{r}textual content{a}textual content{g}textual content{e}:textual content{p}textual content{o}textual content{o}textual content{l}textual content{i}textual content{n}textual content{g}left(textual content{L}proper)) means world common pooling of enter characteristic L, (:textual content{M}textual content{a}textual content{x}textual content{i}textual content{m}textual content{u}textual content{m}:textual content{p}textual content{o}textual content{o}textual content{l}left(textual content{L}proper)) represents world most pooling of enter characteristic L.

Fig. 2
figure 2

Operation steps of spatial consideration residual module.

The appliance of spatial consideration mechanism can enhance the mannequin’s consideration to essential spatial info. The spatial consideration module is added to every node of the ResNet-101 structure to acquire a brand new spatial consideration residual module (Fig. 3). The particular operation steps are as follows: The enter options are processed utilizing max pooling and world common pooling to acquire two channel descriptions. The 2 channel descriptions are collectively utilized to rework the double-layer characteristic map right into a single-layer characteristic map. After a 5 × 5 convolutional operation, the Sigmoid activation perform is used to acquire the weight coefficients.

Fig. 3
figure 3

Operation steps of spatial consideration residual module.

As a consequence of the serial construction having the stacking impact of extra nonlinear activation features, higher non-residual blocks may be obtained. Subsequently, in the mannequin, the spatial consideration module and the channel consideration module exist in a serial kind. This way can concurrently calculate the info of completely different channels in the characteristic map in addition to the native spatial info of every channel, thereby enhancing the mannequin’s capacity to study picture options. Thus, an improved ResNet-101 mannequin primarily based on spatial consideration and channel consideration mechanisms was accomplished.

Analysis indicators of the mannequin

AC, PR, RE, and F1 had been used to guage the diagnostic efficiency of the mannequin. Conventional ResNet-101, SA-ResNet-101, CA-ResNet-101 fashions had been launched. Utilizing ROC curve, evaluation of mannequin on the prognosis of sufferers with cervical cancer threat and prediction efficiency was carried out.

$${textual content{AC}} = left( {{textual content{TP}} + {textual content{TN}}} proper)/{textual content{whole quantity of sufferers}}$$

(3)

$${textual content{PR}} = {textual content{TP}}/left( {{textual content{TP}} + {textual content{FP}}} proper)$$

(4)

$${textual content{RE}} = {textual content{TP}}/left( {{textual content{TP}} + {textual content{FN}}} proper)$$

(5)

$${textual content{F}}1 = 2 occasions {textual content{TP}} occasions {textual content{RE}}/left( {{textual content{PR}} + {textual content{RE}}} proper)$$

(6)

TP denotes true constructive, FN denotes false destructive, TN denotes true destructive, and FP denotes false constructive.

Datasets assortment

Retrospectively, 210 sufferers with EC who underwent pelvic MRI examinations at the imaging heart of XXX Hospital from January 2021 to Might 2024 had been included as research samples. Amongst them, 140 circumstances had been used as the take a look at set, and 70 circumstances as the validation set. All sufferers had been pathologically confirmed, and info similar to primary affected person information (Desk 1), imaging footage, and postoperative recurrence was collected, taking whether or not the affected person has a recurrence after surgical procedure as the endpoint occasion. Based on the ESMO-ESTRO-ESP tips, sufferers had been divided into low-risk EC and high-risk EC.

The sufferers’ MRI photos had been transported to the workstation, the place picture processing algorithms had been utilized to phase and reconstruct the photos.

Inclusion standards: (1) Sufferers recognized with EC; (2) Age between 30 and 64 years; (3) Sufferers who underwent pelvic MRI inside the specified time and had full imaging information; (4) Full medical info, together with primary information, imaging footage, postoperative recurrence standing, and many others.

Exclusion standards: (1) Sufferers with different malignant tumors; (2) Substandard MRI picture high quality (similar to movement artifacts, extreme noise); (3) Incomplete follow-up information.

Desk 1 Basic info of sufferers in the take a look at set and validation set.

Experimental surroundings

The experiments had been primarily performed beneath the TensorFlow deep studying framework, accelerated by GPU. The fashions and coaching code had been written in Python model 3.6, with the built-in improvement surroundings being PyCharm. The configuration was as follows: the graphics card was NVIDIA GeForce RTX 2080 Ti, with 64GB of reminiscence, a central processor of AMD Ryzen Threadripper 2950X, and the system was Home windows 10.

Parameters setting

The mannequin parameters had been as follows: the convolutional kernel measurement was 5 × 5, the convolution operation step was 1, the filter measurement was 5 × 5, the depth was set to 101 layers, the scaling issue was 16, the regularization parameter was 2, and the preliminary studying fee was 0.01.

Statistical processing

SPSPS 22.0 statistical software program was employed. Quantitative information conforming to regular distribution had been introduced as imply ± sd ((:overline{textual content{x}}) ± s), quantitative information that didn’t conform to regular distribution had been expressed utilizing the median and interquartile vary, and categorical information had been expressed utilizing frequency and proportion (%). Non-normally distributed quantitative information had been analyzed by Mann-Whitney take a look at, usually distributed quantitative information by one-way ANOVA, and categorical information by chi-square take a look at. The diagnostic efficiency of every mannequin was assessed by plotting the ROC curves, and the AUC was calculated to check the Sen and Spe of completely different fashions. A two-tailed take a look at with P < 0.05 was thought-about the commonplace for statistical significance. The statistical evaluation outcomes indicated that the improved deep studying mannequin demonstrated vital statistical benefits in the diagnosis of high-risk EC sufferers and in the prediction of postoperative recurrence.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *