Optical coherence tomography real-time feedback control system based on neuromorphic computing

Abstract

The application relates to the technical field of laser precision machining, in particular to an optical coherence tomography real-time feedback control system based on neuromorphic computing, which comprises an optical coherence tomography device and a processing module; the optical coherence tomography device comprises a collecting module and an executing module; the collecting module is used for collecting an analog voltage signal corresponding to an original optical interference signal of a biological tissue; the processing module is used for converting the analog voltage signal into a digital signal stream, inputting the digital signal stream into a neural network model, outputting a target depth position index from the neural network model, and generating a control signal according to the target depth position index; and the executing module is used for controlling the energy output of a surgical laser guided by the optical coherence tomography according to the control signal, simulating a reflex arc mechanism of a biological nervous system, constructing a bionic optical reflex arc architecture, and being low in calculation delay and small in computing power consumption. Therefore, the problems that related technologies cannot match microsecond laser pulses and cannot realize precise control are solved.

Description

Technical Field

[0001] This application relates to the field of laser precision machining technology, and in particular to a real-time feedback control system for optical coherence tomography based on neuromorphic computing. Background Technology

[0002] OCT (Optical Coherence Tomography)-guided femtosecond / picosecond laser surgery systems represent a cutting-edge direction in minimally invasive ophthalmic treatment. However, the current serial processing mode based on the von Neumann architecture suffers from a serious bottleneck: its millisecond-level image reconstruction and feedback delays cannot match the microsecond-level pulse intervals of megahertz lasers, preventing the system from achieving pulse-by-pulse control. This lag can easily lead to laser overcutting, damaging choroidal vessels and causing massive retinal hemorrhage, thus violating clinical safety red lines.

[0003] Furthermore, threshold-based depth recognition algorithms in related technologies also struggle to adapt to the complex optical characteristics of the fundus and individual patient differences, easily leading to misjudgments due to strong reflection interference. When facing micron-level surgeries requiring precise protection of highly complex "sandwich structure" tissues, such as subretinal drug delivery or SRT (Selective Retina Therapy), the lack of submicrosecond response and adaptive recognition capabilities in these technologies makes it difficult to meet the high-precision requirement of "instant cutting and stopping," thus limiting the safety and success rate of the surgery. Summary of the Invention

[0004] This application provides a real-time feedback control system for optical coherence tomography based on neuromorphic computing to solve the problem that related technologies cannot match microsecond-level laser pulses and thus cannot achieve precise control.

[0005] The first aspect of this application provides a real-time feedback control system for optical coherence tomography (OCT) scans based on neuromorphic computing, comprising: an OCT scan device and a processing module. The OCT scan device includes an acquisition module and an execution module. The acquisition module is used to acquire analog voltage signals corresponding to the original optical interference signals of biological tissues. The processing module is used to convert the analog voltage signals into digital signal streams, input the digital signal streams into a neural network model, output a target depth position index from the neural network model, and generate a control signal based on the target depth position index. The neural network model obtains the target depth position index through point-by-point streaming inference of the digital signal stream. The execution module is used to control the energy output of a surgical laser guided by the OCT scan according to the control signal.

[0006] Optionally, the acquisition module includes an optical coherence tomography (OCT) probe and an OCT detector. The OCT probe and the OCT detector are optically linked. The OCT probe acquires the raw optical interference signal of the biological tissue, and the OCT detector outputs an analog voltage signal.

[0007] Optionally, the execution module includes a laser control gate and a laser probe. The laser control gate is an acousto-optic modulator or a seed source trigger interface, which is directly controlled by the control signal output by the processing module.

[0008] Optionally, the processing module includes an analog-to-digital converter (ADC) module and a field-programmable gate array (FPGA) module. The ADC module converts analog voltage signals into digital signal streams. A hardware direct connection data path is established between the ADC module and the FPGA module. The hardware direct connection data path directly inputs the digital signal stream to the FPGA module in a streaming manner without going through general-purpose computer memory and operating system scheduling.

[0009] Optionally, the neural network model is set in the field-programmable gate array module. The neural network model adopts a unidirectional streaming architecture, including a wide convolutional preprocessing module, a multi-layer recursive spiking neuron module, and an output decoding module connected in sequence. The wide convolutional preprocessing module is configured with a one-dimensional convolutional layer for extracting local texture features of the input signal. The multi-layer recursive spiking neuron module contains at least two spiking neuron layers, and residual connections are set between the spiking neuron layers. The multi-layer recursive spiking neuron module is configured to directly superimpose the input current of the shallow spiking neuron layer to the input of the deep spiking neuron layer. The multi-layer recursive spiking neuron module is also configured with recursive connections for transmitting the pulse state at the current moment back to the input of the next moment. The output decoding module is used to output the target depth position index.

[0010] A second aspect of this application provides a real-time feedback control method for optical coherence tomography (OCT) based on neuromorphic computing. The method is applied to the processing component of the real-time feedback control system for OCT based on neuromorphic computing in the first aspect, and includes the following steps: acquiring an analog voltage signal corresponding to the original optical interference signal of biological tissue; converting the analog voltage signal into a digital signal stream; inputting the digital signal stream into a neural network model; the neural network model outputting a target depth position index; generating a control signal based on the target depth position index; wherein the neural network model obtains the target depth position index through point-by-point streaming inference of the digital signal stream; and using the control signal to control the energy output of a surgical laser guided by OCT.

[0011] Optionally, before inputting the digital signal stream into the neural network model, the method further includes: performing resampling or linear interpolation on the digital signal stream to correct the nonlinearity of the wavenumber space; performing a fast Fourier transform on the corrected digital signal stream to convert it into an amplitude scanning signal stream; performing real-time surface localization on the amplitude scanning signal stream to obtain the surface position, determining the surface depth position index of the biological tissue, and aligning the amplitude scanning signal stream based on the surface position; and inputting the aligned amplitude scanning signal stream point by point into the neural network model.

[0012] Optionally, the neural network model outputs a target depth location index, including: using a one-dimensional convolutional layer to extract features from the input current data point and its neighborhood data, and outputting simulated feature values; using spiking neurons to encode the simulated feature values ​​into pulse signals; using recursive connections to combine the membrane potential state of the previous time step for temporal memory updates; and using an output decoding layer to map the pulse features into the confidence voltage value corresponding to the current depth in real time. When the voltage value exceeds the activation threshold, the current time step is marked as the target depth location index.

[0013] Optionally, before inputting the digital signal stream into the neural network model, the method further includes: obtaining a training dataset for the neural network model, wherein the training dataset includes training samples and ground truth labels, the training samples being amplitude scan signal streams and the ground truth labels being position labels corresponding to the amplitude scan signal streams; training the neural network model using the training samples in the training dataset, during which the training loss is calculated using the ground truth labels and the predicted target depth position index of the output of the neural network model, and the model parameters of the neural network model are updated based on the training loss; during the training process, a cut-off depth is randomly generated, data of the length corresponding to the cut-off depth at the front end of the amplitude scan signal stream in the training samples is removed, and the remaining data of the amplitude scan signal stream is shifted forward to align, while the position labels of the training samples are adjusted synchronously.

[0014] Optionally, generating a control signal based on the target depth location index includes: receiving the biological tissue surface location index; calculating the difference between the biological tissue surface location index and the target depth location index; calculating the remaining tissue thickness based on the difference and physical resolution; comparing the remaining tissue thickness with a preset safety threshold; obtaining a quantity threshold; and generating a control signal when it is identified that the remaining tissue thickness of multiple consecutive sampling frames is less than the safety threshold and the number of multiple sampling frames is greater than the quantity threshold.

[0015] Therefore, this application has the following beneficial effects: This application embodiment can acquire analog voltage signals corresponding to the original optical interference signals of biological tissue through an acquisition module; the processing module converts the analog voltage signals into digital signal streams, inputs the digital signal streams into a neural network model, and the neural network model outputs a target depth position index. A control signal is generated based on the target depth position index. The execution module controls the energy output of the surgical laser guided by optical coherence tomography (OCT) according to the control signal, simulating the reflex arc mechanism of the biological nervous system and constructing a biomimetic optical reflex arc architecture. This architecture features low computational latency, low computational power consumption, and strong anti-interference capabilities, meeting the real-time requirements of millisecond-level or even microsecond-level laser surgery. Therefore, it solves the problem that related technologies cannot match microsecond-level laser pulses and cannot achieve precise control.

[0016] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0017] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a real-time feedback control system for optical coherence tomography based on neuromorphic computing, according to an embodiment of this application. Figure 2 This is a block diagram of a streaming spiking neural network model according to an embodiment of this application; Figure 3 This is a schematic diagram of a biomimetic optical reflection arc architecture according to an embodiment of this application; Figure 4 Here is a flowchart of a threshold selection algorithm according to an embodiment of this application; Figure 5 This is a flowchart illustrating the deployment of a field-programmable gate array (FPGA) module for a streaming convolutional spiking neural network model according to an embodiment of this application. Figure 6 This is a flowchart of a real-time feedback control method for optical coherence tomography based on neuromorphic computing according to an embodiment of this application; Figure 7 This is a flowchart illustrating the real-time feedback control method for optical coherence tomography based on neuromorphic computing provided in an embodiment of this application. Detailed Implementation

[0018] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0019] Among related technologies, OCT-guided intelligent laser surgery systems, which combine OCT imaging with femtosecond / picosecond laser treatment, represent a cutting-edge development in minimally invasive ophthalmic surgery. These systems aim to utilize the depth resolution capabilities of OCT to monitor the laser ablation depth in real time, achieving precise removal of lesions while protecting surrounding healthy tissue. Current OCT-guided laser surgery systems typically employ a traditional computational processing flow based on the von Neumann architecture, with a typical operating mode following a serial processing path of "signal acquisition, image reconstruction, image processing, and feedback control."

[0020] While related technologies have shown excellent performance in offline diagnosis, they have revealed significant shortcomings in real-time feedback control for high-frequency laser surgery. Current image reconstruction and post-processing algorithms typically involve extensive data transfer and floating-point operations, with processing times often on the order of milliseconds or even tens of milliseconds. However, the femtosecond or picosecond lasers used in modern ophthalmic surgery usually operate at repetition frequencies in the megahertz range, meaning pulse intervals are only on the order of microseconds or even nanoseconds. This huge time difference between "millisecond-level feedback" and "microsecond-level pulses" makes it impossible for the control system to achieve true "pulse-by-pulse" control. Due to the aforementioned computational delay, by the time the system determines that the laser has reached the target depth through image recognition, the laser has actually already emitted thousands of subsequent pulses. This feedback lag effect can easily cause the laser to cut through the target layer, thereby damaging the deep choroidal vessels. Clinically, this is a "safety red line" to prevent massive retinal hemorrhage; once breached, it will cause irreversible vision damage. Existing depth recognition largely relies on threshold-based image segmentation algorithms. However, the optical properties of fundus tissues are complex. For example, surface tissues such as the internal limiting membrane have strong reflective properties, making them easily misidentified as target layers by traditional algorithms, leading to premature laser termination. In addition, there are significant individual differences in the intensity of tissue scattering signals among different patients, making it difficult for traditional algorithms with fixed parameters to achieve robust generalization among different patients, thus limiting the success rate of the surgery.

[0021] In specific minimally invasive retinal surgery scenarios, the control precision of related technologies can no longer meet the increasingly sophisticated clinical needs. In surgeries constructing subretinal drug delivery channels, the laser must precisely penetrate all upper neural tissues, including the nerve fiber layer and ganglion cell layer, and "just right" stop at the surface of the retinal pigment epithelium (RPE) to form the drug delivery channel. The control precision of related technologies is insufficient to achieve this "cut-and-stop" at the micrometer level, easily leading to penetration of the RPE or failure to establish the channel. Selective retinal therapy targets a highly challenging "sandwich structure": the upper layer contains invaluable photoreceptor cells, thermal damage of which can cause blindness; the lower layer contains choroidal vessels, puncture of which can cause bleeding; the target is only the middle RPE layer. Because existing image reconstruction-based control methods lack sub-microsecond response speeds, it is difficult to ensure the absolute safety of adjacent tissues while destroying RPE cells.

[0022] To address the shortcomings of related technologies, this application proposes a real-time feedback control system for optical coherence tomography based on neuromorphic computing, which solves the problem that related technologies cannot match microsecond-level laser pulses and thus cannot achieve precise control.

[0023] The following describes an embodiment of the real-time feedback control system for optical coherence tomography (OCT) based on neuromorphic computing, with reference to the accompanying drawings. Addressing the problem mentioned in the background art that related technologies cannot match microsecond-level laser pulses and thus cannot achieve precise control, this application provides a real-time feedback control system for OCT based on neuromorphic computing. In this system, an acquisition module acquires analog voltage signals corresponding to the original optical interference signals of biological tissues; a processing module converts the analog voltage signals into digital signal streams, which are then input into a neural network model. The neural network model outputs a target depth position index, and a control signal is generated based on this index. An execution module controls the energy output of the surgical laser guided by OCT according to the control signal, simulating the reflex arc mechanism of a biological nervous system and constructing a biomimetic optical reflex arc architecture. This architecture features low computational latency, low computational power consumption, and strong anti-interference capabilities, meeting the real-time requirements of millisecond-level or even microsecond-level laser surgery. Thus, it solves the problem that related technologies cannot match microsecond-level laser pulses and thus cannot achieve precise control.

[0024] Specifically, Figure 1 This is a flowchart illustrating a real-time feedback control system for optical coherence tomography based on neuromorphic computing, provided in an embodiment of this application.

[0025] like Figure 1As shown, the real-time feedback control system for optical coherence tomography based on neuromorphic computing includes: an optical coherence tomography device 101 and a processing module 102. The optical coherence tomography device 101 includes an acquisition module 1011 and an execution module 1012.

[0026] The acquisition module 1011 is used to acquire the analog voltage signal corresponding to the original optical interference signal of the biological tissue; the processing module 102 is used to convert the analog voltage signal into a digital signal stream, input the digital signal stream into a neural network model, output the target depth position index from the neural network model, and generate a control signal based on the target depth position index. The neural network model obtains the target depth position index through point-by-point streaming inference of the digital signal stream; the execution module 1012 is used to control the energy output of the surgical laser guided by optical coherence tomography according to the control signal.

[0027] Point-by-point streaming inference is a data processing mode that does not require waiting for a complete frame of data to be acquired. Instead, it performs feature extraction and state updates immediately with each input data point, following the sampling clock, achieving real-time processing that is calculated as it is input. Point-by-point streaming inference can reduce latency from milliseconds to microseconds or even lower. Biological tissues refer to the basic structural units that make up an organism (such as skin, retina, blood vessels, etc.). In optical imaging, biological tissues have the characteristics of scattering and reflecting light, and tissue structures at different depths will reflect light of different intensities.

[0028] It is understood that the embodiments of this application construct a real-time feedback control system for optical coherence tomography with a biomimetic reflective arc architecture through an acquisition module 1011, an execution module 1012, and a processing module 102. The acquisition module 1011 directly acquires the original optical interference analog voltage signal of biological tissue and inputs the analog voltage signal to the processing module 102. The processing module 102 converts the analog voltage signal into a digital signal stream and extracts features from the digital signal stream in real time and outputs the target depth position index through a neural network model and a point-by-point streaming inference method. Then, a control signal is generated based on the target depth position index. The execution module 1012 receives the control signal and physically cuts off the energy output of the surgical laser guided by optical coherence tomography within microseconds, thereby reducing computational latency and computing power consumption, breaking through the timing bottleneck, and realizing pulse-by-pulse closed-loop control.

[0029] In this embodiment, the acquisition module 1011 includes an optical coherence tomography (OCT) probe and an OCT detector. The OCT probe and the OCT detector are optically linked. The OCT probe acquires the raw optical interference signal of biological tissue, and the OCT detector outputs an analog voltage signal.

[0030] Among them, the optical coherence tomography (OCT) probe is an optical sensor device, which typically includes components such as fiber optic couplers and scanning mirrors. Its function is to emit broadband near-infrared light into biological tissue and collect reference light and sample light backscattered from different depth layers of the tissue to form a raw optical interference signal carrying depth information. The OCT detector is a high-sensitivity, low-noise photoelectric conversion device used to receive two interference light signals from the probe. It cancels common-mode noise through differential amplification technology and linearly converts the weak optical interference signal into a high signal-to-noise ratio analog voltage signal to prepare for subsequent analog-to-digital conversion. The optical link refers to the optical transmission path between the probe and the detector, which is usually composed of single-mode fiber or free-space optical path and is used to transmit the interference light signal without loss.

[0031] For example, the optical link of the acquisition module 1011 adopts a fiber-optic Michelson interferometer or Mach-Zehnder interferometer architecture. The optical coherence tomography probe contains a scanning galvanometer and a focusing lens group. The swept laser emitted by the optical coherence tomography light source is split into reference light and sample light by the fiber optic coupler. The sample light is focused into the biological tissue by the optical coherence tomography probe. The backscattered light returns and interferes with the reference light in the coupler. The optical coherence tomography detector is preferably a balanced photodetector, which receives the interference light signal, suppresses common-mode noise through a differential amplifier circuit, and outputs a high-frequency analog radio frequency voltage signal containing depth information. The frequency range of this analog signal is usually adapted to the sweep speed of the light source.

[0032] It is understood that the acquisition module 1011 in this embodiment of the application consists of an optical coherence tomography probe and an optical coherence tomography detector. The two transmit optical signals through an optical link. The optical coherence tomography probe is responsible for acquiring the raw optical interference signal containing depth information of biological tissue and transmitting the raw optical interference signal to the optical coherence tomography detector via the optical link. Finally, the optical coherence tomography detector completes the photoelectric conversion and outputs an analog voltage signal for use by the subsequent processing module.

[0033] In this embodiment, the execution module 1012 includes a laser control gate and a laser probe. The laser control gate is an acousto-optic modulator or a seed source trigger interface, which is directly controlled by the control signal output by the processing module 102.

[0034] Among them, the control signal is a level control signal; the laser control gate is a high-speed switching device located in the laser optical path or driving circuit. In this embodiment, it is an AOM (Acousto-Optic Modulator, which quickly deflects the beam through acoustic diffraction to cut off the optical path) or a seed source trigger interface (which directly controls the electrical pulse output of the laser seed). The core characteristic of both is that the response speed is extremely fast and can directly respond to the control signal issued by the processing module 102.

[0035] It is understood that the execution module 1012 in this embodiment consists of a laser control gate and a laser probe. The laser control gate specifically uses an acousto-optic modulator or a seed source trigger interface as a high-speed switching element. This element skips any intermediate buffer or operating system and is directly controlled by the physical level control signal output in real time by the processing module 102. When the processing module 102 issues a control signal, it immediately drives the laser control gate to cut off the energy output of the surgical laser, thereby preventing the laser energy from continuing to act on the biological tissue through the laser probe, thus achieving precise instantaneous shutdown protection.

[0036] In this embodiment, the processing module 102 includes an analog-to-digital converter module and a field-programmable gate array (FPGA) module. The analog-to-digital converter module converts analog voltage signals into digital signal streams. A hardware direct connection data path is established between the analog-to-digital converter module and the FPGA module. The hardware direct connection data path does not go through general-purpose computer memory and operating system scheduling, and directly inputs the digital signal stream to the FPGA module in a streaming manner.

[0037] The analog-to-digital converter (ADC) module, also known as a high-speed analog-to-digital converter, discretizes continuous analog voltage signals at an extremely high sampling rate, converting them into a binary digital signal stream that can be processed by a computer. The ADC module is connected to the output of the optical coherence tomography (OCT) detector via an RF cable. The field-programmable gate array (FPGA) module is an integrated circuit chip with highly parallel processing capabilities and hardware reconfigurability. It can directly run streaming neural network models and perform real-time inference on the input digital signals without relying on the instruction cycles of a general-purpose CPU (Central Processing Unit), resulting in extremely low processing latency. The hardware direct-connect data path refers to a dedicated physical transmission channel connecting the ADC module and the FPGA module. In this embodiment, a high-speed serial interface or a parallel differential interface is used for connection. This path directly connects the two at the physical level, without going through the main memory of a general-purpose computer or the kernel scheduling of the operating system. This eliminates data copying, context switching, and interrupt latency, ensuring that data enters the FPGA module in a zero-buffered streaming manner as it is acquired and transmitted. This is the core architecture for achieving deterministic low latency.

[0038] It is understood that the processing module 102 in this embodiment is composed of an analog-to-digital converter module and a field-programmable gate array (FPGA) module working together. The analog-to-digital converter module is responsible for quantizing the input analog voltage signal into a digital signal stream in real time. To ensure ultimate real-time performance, a dedicated hardware direct connection data path is constructed between the analog-to-digital converter module and the FPGA module. This path completely bypasses the scheduling mechanism of general-purpose computer memory and operating system, avoids data caching and interruption delays in traditional architectures, and directly inputs the digital signal stream to the FPGA module in a streaming manner without obstruction for immediate processing, thereby ensuring the system's rapid response capability to changes in the state of biological tissues.

[0039] In this embodiment, the neural network model is housed within a field-programmable gate array (FPGA) module. The neural network model employs a unidirectional streaming architecture, comprising a wide convolutional preprocessing module, a multi-layer recursive spiking neuron module, and an output decoding module connected sequentially. The wide convolutional preprocessing module is configured with a one-dimensional convolutional layer for extracting local texture features of the input signal. The multi-layer recursive spiking neuron module contains at least two spiking neuron layers, with residual connections between the spiking neuron layers. The multi-layer recursive spiking neuron module is configured to directly superimpose the input current of the shallow spiking neuron layer onto the input of the deep spiking neuron layer. The multi-layer recursive spiking neuron module also has recursive connections for transmitting the pulse state at the current moment back to the input at the next moment. The output decoding module is used to output the target depth position index.

[0040] The unidirectional streaming architecture is a data flow mode where signals are processed sequentially in a single direction (from input to output) without waiting for future data or performing global backtracking. This architecture aligns with the pipeline characteristics of field-programmable gate array (FPGA) modules, ensuring both real-time processing and determinism. The wide convolutional preprocessing module is the first-layer feature extraction unit of the neural network model. Wide convolution refers to the use of larger convolutional kernels, designed to capture contextual information or macroscopic texture features over a wider range in the input signal. It is configured with one-dimensional convolutional units, batch normalization units, and activation units. After convolution, the convolution result is processed by the batch normalization unit to stabilize the data distribution. Subsequently, nonlinear features are introduced through the activation function in the activation unit, outputting an analog voltage feature stream. The one-dimensional convolutional layer is specifically designed for processing... Convolution operations on sequential data (such as time-series signals) extract local texture features (such as specific shapes, edges, or frequency components of waveforms) by sliding along the time axis, transforming the original voltage signal into a high-dimensional feature representation. Residual connections are a cross-layer connection mechanism that allows the input current from shallow layers to skip intermediate layers and be superimposed on the input of deeper layers, solving the gradient vanishing problem in deep networks, ensuring that shallow features can be effectively transmitted to deeper layers, accelerating convergence and improving accuracy. Recursive connections are a time-dimensional feedback mechanism that feeds back the neuron spiking state at the current moment to the input of the next moment, enabling the neural network model to have memory capabilities and capture the dynamic evolution of the signal along the time axis. The multi-layer recursive spiking neuron module is the core computational part of the network, simulating the "firing-integration" mechanism of biological neurons, specifically: The multi-layer recursive spiking neuron module adopts a deep cascaded architecture, which contains three main neuron layers and integrates parametric spiking neurons (Parametric LIF, P=τ, where ParametricLIF and ParametricLeaky Integrate-and-Fire are configurable parameter sets for the leaky integral firing model; time constant τ), recursive connections, and residual connections.

[0041] Specifically, for example, such as Figure 2 As shown, the first layer of the multi-layer recurrent spiking neuron module is the encoding layer, containing LIF (Leaky Integrate-and-Fire) 1 neuron nodes and a fully connected Linear (64->128) layer. LIF 1 receives a 64-dimensional analog signal from the preprocessing module, encodes it into a sparse pulse sequence using the parameterized membrane potential time constant (P=τ), and then expands the feature dimension to 128 dimensions through the fully connected layer; the second layer is the core recurrent layer, containing addition nodes (⊕), LIF 2 neuron nodes, and a fully connected Linear (128->128) layer, as shown. Figure 2The dashed arrow below LIF 2 indicates a recursive connection. The pulse state output by LIF 2 is fed back to the adder node at its input. This autoregressive mechanism enables the network to remember the organizational structure features of the previous moment, thereby assisting in the judgment at the current moment. The third layer is the residual enhancement layer, which contains the adder node (⊕) and LIF 3 neuron nodes. Figure 2 The dashed line at the top center, crossing the arrow, indicates the residual connection. This embodiment also introduces an identity mapping path, which directly passes the characteristic current output from the first-layer Linear (i.e., the input of the second layer) across the second layer and superimposes it onto the adder node at the input of the third layer. This residual connection constructs a direct gradient transport channel, effectively solving the gradient vanishing problem in streaming training of deep SNNs (Spiking Neural Networks). This allows the model to retain the original feature amplitude while further extracting deep temporal features through LIF 3. LIF 3 is also configured with recursive connections, further enhancing the ability to capture long-distance temporal dependencies.

[0042] It should be understood that in other embodiments of this application, the neural network model may also be trained to directly output the encoded remaining thickness value, without any specific limitation.

[0043] It is understood that the neural network model deployed in the field-programmable gate array module in this application embodiment adopts an efficient unidirectional streaming architecture, consisting of a wide convolutional preprocessing module, a multi-layer recursive spiking neuron module, and an output decoding module connected in sequence. The wide convolutional preprocessing module has a built-in one-dimensional convolutional layer, which is specifically used to extract key local texture features from the input signal. The subsequent multi-layer recursive spiking neuron module contains at least two spiking neuron layers, and the input current of the shallow layer is directly superimposed to the input end of the deep layer through residual connections to retain feature information. At the same time, the recursive connection is used to backpropagate the pulse state at the current moment to the next moment, giving the network the ability to remember and process the temporal dynamic characteristics. Finally, the output decoding module analyzes the spiking activity of the neurons and accurately outputs the target depth position index indicating the tissue location, realizing end-to-end real-time mapping from the original signal to depth control.

[0044] The real-time feedback control system for optical coherence tomography (OCT) based on neuromorphic computing proposed in this application can acquire analog voltage signals corresponding to the original optical interference signals of biological tissues through an acquisition module; convert the analog voltage signals into digital signal streams through a processing module, input the digital signal streams into a neural network model, output a target depth position index from the neural network model, and generate control signals based on the target depth position index; and use an execution module to control the energy output of the surgical laser guided by OCT according to the control signals, thereby simulating the reflex arc mechanism of the biological nervous system and constructing a biomimetic optical reflex arc architecture. This system features low computational latency, low computational power consumption, and strong anti-interference capabilities, meeting the real-time requirements of millisecond-level or even microsecond-level laser surgery.

[0045] The following specific embodiment further describes the real-time feedback control system for optical coherence tomography based on neuromorphic computing.

[0046] Figure 3 This is a schematic diagram of a biomimetic optical reflection arc architecture provided in this embodiment.

[0047] like Figure 3 As shown, this embodiment proposes a real-time feedback control system for optical coherence tomography based on neuromorphic computing. In its architecture design, this system abandons the high-latency architecture of traditional "computer acquisition + software image reconstruction + manual / algorithm judgment," instead simulating the "receptor-center-effector" reflex arc mechanism of the biological nervous system, constructing a direct hardware connection from signal acquisition to laser control.

[0048] This embodiment constructs a biomimetic optical reflection arc architecture, which sequentially includes: an acquisition module 1011, a processing module 102, and an execution module 1012. The acquisition module 1011 includes an optical coherence tomography (OCT) probe and an OCT detector; the processing module 102 includes an analog-to-digital converter (ADC) module and a field-programmable gate array (FPGA) module; the execution module 1012 includes a laser control gate and a laser probe; the FPGA module internally includes a preprocessing algorithm, a streaming convolutional spiking neural network (SPNN) model, and a threshold selection algorithm. This system adopts a "de-image processing" mode, directly processing the A-scan (amplitude scan signal stream) signal without requiring B-scan (section scan signal stream) image reconstruction. The FPGA module relies on the streaming convolutional spiking neural network model to perform inference on the preprocessed A-scan, achieving on-demand computation of the data. When the streaming spiking neural network model detects that the remaining thickness is below the safety threshold, the system directly triggers the laser control gate through the GPIO (General-Purpose Input / Output) pin, forming a closed-loop hard real-time control from optical signal input to optical energy blocking, thereby meeting the requirements of laser ablation surgery for extreme safety and response speed.

[0049] The acquisition module 1011, acting as the system's "visual nerve ending," is used to acquire depth interference signals from biological tissue in real time. The acquisition module 1011 includes an optical coherence tomography (OCT) probe and an OCT detector. The OCT probe emits a probe beam towards the tissue to be ablated and collects backscattered light. The OCT detector is optically connected to the OCT probe and receives the interference light signal, converting it into a high-frequency analog voltage signal. This signal is the original one-dimensional A-scan signal, which contains structural information about tissue depth variations.

[0050] Specifically, the optical link of the acquisition module 1011 employs a fiber-optic Michelson interferometer or Mach-Zehnder interferometer architecture. The optical coherence tomography (OCT) probe internally includes a scanning mirror and a focusing lens group. The swept-frequency laser emitted from the OCT source (center wavelength 1310 nm, scan frequency 100 kHz-400 kHz) is split into reference light and sample light via a fiber optic coupler. The sample light is focused onto the biological tissue by the OCT probe, and the backscattered light returns and interferes with the reference light in the coupler. The OCT detector is preferably a balanced photodetector. It receives the interference light signal, suppresses common-mode noise through a differential amplifier circuit, and outputs a high-frequency analog radio frequency voltage signal containing depth information. The frequency range of this analog signal is typically adapted to the sweep speed of the light source, for example, between 10 MHz and 500 MHz.

[0051] Processing module 102, acting as the system's "reflection center," is responsible for high-speed digitization and neuromorphic computation processing of the signal. Processing module 102 includes an analog-to-digital converter (ADC) module and a field-programmable gate array (FPGA) module. The ADC module is connected to the output of the optical coherence tomography (OCT) detector via an RF cable. To match the high-resolution requirements of OCT, the ADC module's sampling rate must satisfy the Nyquist sampling theorem, configured with at least 12-bit precision and a sampling rate of 500 MSPS or higher, used to convert analog voltage signals into digital signal streams. The ADC module and the FPGA module are connected via a high-speed serial interface or a parallel differential interface. The digital signal stream output by the ADC module does not pass through a DDR (Double Data Rate Synchronous Dynamic Random-Access Memory) buffer, but is directly streamed into the FIFO (First-In, First-Out) buffer at the programmable logic level of the FPGA module for direct reading by subsequent algorithm pipelines.

[0052] Specifically, the field-programmable gate array (FPGA) module serves as the core of the system in this embodiment. Its internal logic resources incorporate a parallel processing pipeline to achieve real-time processing of OCT signals at the millisecond or even microsecond level. For example... Figure 3 As shown, the internal logic architecture of the field-programmable gate array module mainly includes: preprocessing algorithm, streaming convolutional spiking neural network model, and threshold selection algorithm.

[0053] In this embodiment, a preprocessing algorithm is used to perform signal conditioning and domain transformation on the raw digital signal stream input from the analog-to-digital conversion module. The preprocessing algorithm is configured with k-clock resampling logic, fast Fourier transform logic, and surface tracking logic.

[0054] Specifically, the original signal is converted into a depth domain A-scan sequence after resampling and fast Fourier transform processing; the surface tracking logic is used to detect strong reflection points in the A-scan sequence in real time where the signal intensity first undergoes a significant change, and marks the point as the index position of the tissue surface.

[0055] Furthermore, the preprocessing algorithm may also include filtering logic for eliminating baseline drift or shift register logic for data alignment, so as to provide standardized input data for subsequent model inference.

[0056] The streaming convolutional spiking neural network model is the core unit for realizing tissue structure recognition.

[0057] In this embodiment, the streaming convolutional spiking neural network model is deployed inside the field-programmable gate array module after quantization processing. It adopts a unidirectional streaming inference architecture, which can process the input A-scan data point by point and output the index position of the target layer in the A-scan sequence in real time.

[0058] It should be understood that in other embodiments of this application, the streaming convolutional spiking neural network model can also be trained to directly output the encoded remaining thickness value, and this application does not impose specific limitations on this.

[0059] The threshold selection algorithm is a key logic unit for ensuring surgical safety. It is configured within the field-programmable gate array (FPGA) module and is used to receive the target depth location index derived in real time, calculate the current remaining tissue thickness, and generate the final laser control signal according to preset rules. The core of these preset rules is a preset safety threshold. This preset safety threshold refers to the minimum biological tissue thickness that must be preserved during laser ablation to protect deep critical tissues from thermal or penetrating damage. It is specifically set according to actual needs and is not specifically limited here.

[0060] In this embodiment, the safety threshold is a configurable digital value whose physical meaning corresponds to a depth at the micrometer level. This threshold is set by the host computer software based on the patient's preoperative diagnostic data before the surgery begins, and is written into the configuration register of the field-programmable gate array module via a bus interface.

[0061] The execution module 1012 acts as the system's "muscle actuator," directly responding to the control commands of the processing module 102 to physically cut off the laser. The execution module 1012 includes a laser control gate and a laser probe. The laser control gate is specifically implemented as an acousto-optic modulator or a seed source trigger interface for the laser. A high-speed GPIO pin of the field-programmable gate array (FPGA) module is connected to the laser control gate via a level conversion circuit. Under normal ablation conditions, the FPGA module outputs a high level (e.g., a 3.3V TTL (Transistor-Transistor Logic) signal), activating the acousto-optic modulator and allowing femtosecond / picosecond laser pulses to pass through and be focused on the probe tip. When the threshold selection algorithm within the FPGA module determines that the stopping condition has been met, the GPIO pin flips to a low level (0V) within nanoseconds, physically cutting off the laser output optical path or stopping the seed source trigger, thereby achieving microsecond-level instantaneous cutting and stopping.

[0062] Figure 2 The block diagram of the streaming spiking neural network model provided in this embodiment.

[0063] like Figure 2The diagram illustrates the architecture of the streaming spiking neural network model in this embodiment, which mainly includes an input module, a wide convolutional preprocessing module, a multi-layer recursive spiking module, and an output module connected in sequence. This model is configured to be deployed in hardware accelerators such as field-programmable gate arrays (FPGAs) for point-by-point streaming processing of OCT A-scan signals.

[0064] The input module receives digitized A-scan data from the analog-to-digital conversion module. Unlike traditional batch processing, this embodiment uses a streaming input mode. Figure 2 As shown in “Point-by-point input of A-scan data 1×1”, in each clock cycle, the network only receives a scalar data point at the current depth position, without requiring the model to infer a complete A-scan data of length 512.

[0065] The wide convolutional preprocessing module, serving as the front end for feature extraction, is equipped with one-dimensional convolutional units, batch normalization units, and activation units. For example... Figure 2 As shown, the one-dimensional convolutional unit (Conv1d) is configured with parameters K=11 and C=64. K=11 indicates a kernel size of 11, meaning that when processing the current point, the model can simultaneously perceive the local context information of 11 adjacent points. This large receptive field design effectively smooths the inherent speckle noise of OCT signals. C=64 indicates that the input one-dimensional scalar signal is mapped to a 64-dimensional feature vector, enriching the signal's representational capabilities. After convolution, the result is processed by a batch normalization unit to stabilize the data distribution, and then nonlinear features are introduced through the LeakyReLU activation function in the activation unit, outputting an analog voltage feature stream.

[0066] Specifically, the wide convolution preprocessing module implements streaming computation within the field-programmable gate array (FPGA) module through a shift register group and parallel multiply-accumulate units. As ADC data is input point by point, the data flows sequentially through the shift registers. In each clock cycle, the 11 data points covered by the convolution kernel (K=11) are read simultaneously, and weighted summation and bias operations are performed in parallel using DSP slices. This hardware architecture ensures that the latency of the convolution operation is constant, eliminating the need to wait for a complete data frame buffer.

[0067] The multi-layer recursive pulse module is the core computing unit of this invention, such as... Figure 2 As shown in the dashed box, it adopts a deep cascaded architecture, which contains three main neuron levels and integrates parametric spiking neurons (P=τ), recursive connections, and residual connections.

[0068] Specifically, the first layer is the encoding layer, containing LIF 1 neuron nodes and a fully connected Linear layer (64->128). LIF 1 receives a 64-dimensional analog signal from the preprocessing module and encodes it into a sparse pulse sequence using a parameterized membrane potential time constant (P=τ). The feature dimension is then expanded to 128 dimensions through a fully connected layer. The second layer is the core recursive layer, containing addition nodes (⊕), LIF 2 neuron nodes, and a fully connected Linear layer (128->128). Figure 2 The dashed arrow below LIF 2 indicates a recurrent connection, where the impulse state output by LIF 2 is fed back to the adder node at its input. This autoregressive mechanism allows the network to "memorize" the organizational structure features of the previous time step, thus aiding in the judgment at the current time step. The third layer is a residual enhancement layer, containing adder nodes (⊕) and LIF 3 neurons. The dashed line crossing the arrow at the top of the diagram indicates a residual connection. This embodiment introduces an identity mapping path, which directly crosses the feature current output of the first layer's Linear layer (i.e., the input of the second layer) and superimposes it onto the adder node at the input of the third layer. This residual connection constructs a direct gradient transport channel, effectively solving the gradient vanishing problem in deep SNNs during streaming training, allowing the model to retain the original feature amplitude while further extracting deep temporal features through LIF 3. LIF 3 is also configured with recurrent connections to further enhance the ability to capture long-distance temporal dependencies.

[0069] Specifically, to more clearly illustrate the execution process of parameterized spiking neurons in LeakyReLU (a modified linear unit with leakage, an activation function for artificial neural networks) digital logic, this embodiment presents its discretized dynamic model. For the l-th layer... One neuron, at time step The membrane potential renewal and pulse delivery follow the following equation: (1) Charging and attenuation process:

[0070] (2) Pulse firing mechanism:

[0071] (3) Membrane potential reset process:

[0072] in, This represents the updated membrane potential at the current moment, which is stored in the status register within the field-programmable gate array module. This represents the membrane potential in the previous clock cycle; This represents the total input current at the current moment. For a hierarchy with recursive connections, this input current includes not only the feedforward input of the previous layer but also the recursive input of the current layer at the previous moment, i.e.:

[0073] This represents the feedforward input vector from the previous level at the current moment. In the hybrid driving mode of this embodiment, this vector can be the pulse sequence output by the previous level or the simulated feature stream after linear transformation. This represents the feedforward synaptic weight matrix, which defines the connection strength between neurons in the previous layer and neurons in the current layer, and corresponds to the static weight parameters in the field-programmable gate array module memory. This represents the output pulse vector of the current layer at the previous time step, which represents the network's state memory at that time step. This represents the recursive synaptic weight matrix, which defines the connection strength of the current layer neurons' own state backpropagation and is used to capture the temporal correlation of OCT signals in the depth direction. The parameterized membrane potential time constant is a learnable parameter that is automatically optimized during training using a backpropagation algorithm. It is used to control the decay rate of the membrane potential to adapt to the temporal characteristics of different levels of the OCT signal. The output is a binary pulse signal (0 or 1); This is the preset pulse firing threshold.

[0074] Through the discretization equations described above, the field-programmable gate array (FPGA) module can use multipliers and adders to complete an iterative update of the neuron state in each clock cycle, thereby achieving microsecond-level streaming inference.

[0075] The fully connected layer is configured to perform feature dimension mapping.

[0076] Specifically, since this embodiment uses a spiking neural network, the input signal... It consists of sparse binary pulses (0 or 1). Therefore, in the hardware implementation of the field-programmable gate array (FPGA) module, the fully connected layer does not perform regular floating-point matrix multiplication operations, but is instead configured as an accumulator array. Specifically, the logic is as follows: when an input port is detected... There is a pulse (i.e.) When this happens, the logic unit will store the corresponding pre-stored synaptic weights. Directly accumulated to the output neuron In the membrane potential register; when the input is 0, no operation is performed.

[0077] This "addition instead of multiplication" operation mechanism avoids consuming the expensive DSP (Digital Signal Processor) multiplier resources in the field-programmable gate array module, significantly reducing the dynamic power consumption of the system and improving the inference speed.

[0078] The output module decodes deep features into physical control signals and includes a dimensionality reduction mapping block and a point-by-point output block. The dimensionality reduction mapping block contains a fully connected Linear (128->1) layer, which aggregates and maps 128-dimensional deep pulse features into one-dimensional analog voltage values. The point-by-point output block, as shown in the figure ("Point-by-point output voltage value 1×1"), outputs the confidence voltage curve corresponding to the current depth in real time as the input progresses. When the curve shows a peak or exceeds a preset threshold, it indicates that the laser focus has reached the target layer position.

[0079] This embodiment constructs the streaming convolutional spiking neural network model architecture shown in the figure above in a field-programmable gate array module. It utilizes wide convolution for noise reduction, recursive connections to remember depth information, and residual connections to ensure training stability, thereby achieving high-precision, low-latency streaming processing of OCT signals.

[0080] Figure 4 The flowchart of the threshold selection algorithm provided in this embodiment is shown.

[0081] like Figure 4 The diagram shows a flowchart of the threshold selection algorithm in this embodiment. This algorithm is configured within the field-programmable gate array (FPGA) module to analyze and process the results in real time and generate the final laser control command. Its core lies in implementing a highly reliable and secure decision-making mechanism that includes multi-frame sliding window verification and state latching.

[0082] Specifically, the threshold selection algorithm first receives the biological tissue surface location index from the preprocessing algorithm. and target depth location index from neural network model algorithm Subsequently, the algorithm's internal arithmetic logic unit calculates the remaining tissue thickness corresponding to the current A-scan in real time according to the following formula. :

[0083] in, This represents the physical resolution of a single pixel in the axial direction for the OCT system.

[0084] It should be noted that if the neural network model is directly trained to output the remaining thickness value, the threshold selection algorithm will directly use the model's output value as... The above calculations are not required.

[0085] After calculating the remaining thickness, the algorithm enters the initial comparison stage. For example... Figure 4 As shown in the first diamond-shaped decision box, it will With the preset safety threshold Compare. If If the current state is determined to be "safe," the system executes the safety branch: the GPIO pin continues to output a high level (laser on), and the shift register used for verification is cleared or reset, preparing for the next A-scan judgment; if If the current state is determined to be "potentially stopped", the algorithm proceeds to the second-level verification process. In the second-level verification process, the algorithm shifts the "potentially stopped" result into a shift register of length N, which stores the judgment results of the most recent N A-scan frames.

[0086] Subsequently, the algorithm enters the final verification stage, such as... Figure 4 As shown in the second diamond-shaped decision box, the system checks whether all N results in the shift register are "potentially stopped". If not (i.e., at least one frame in the most recent N frames is safe), it is determined to be occasional noise or transient fluctuation, and the system returns to the safe branch, continuing to keep the laser on; if yes (i.e., the thickness is detected to be below the threshold for N consecutive frames), the algorithm confirms that the stopping condition has been met and executes the stop branch. In the stop branch, the GPIO pin is immediately toggled low, physically cutting off the laser control gate and stopping the laser from emitting.

[0087] In particular, such as Figure 4 As shown at the end of the process flow, this stop state has a latching characteristic. Once the GPIO output goes low, the logic state machine inside the field-programmable gate array (FPGA) module enters a locked state and no longer responds to subsequent A-scan signal changes. This state can only be released upon receiving a clear manual reset command from the doctor or the host computer system. This design ensures that the laser is permanently cut off when the surgery reaches a critical point, preventing signal fluctuations caused by subsequent tissue morphology changes and thus avoiding accidental laser reconnection, thereby greatly improving safety.

[0088] Figure 5 This embodiment provides a flowchart of the deployment process for the field-programmable gate array module of the streaming convolutional spiking neural network model.

[0089] To efficiently run the trained streaming convolutional spiking neural network in a field-programmable gate array (FPGA) module, this embodiment also provides a specific hardware deployment method. For example... Figure 5 As shown, the deployment process includes the following key steps: Step S11: Model Quantization and Perception Training. Because the internal logic resources of the field-programmable gate array (FPGA) module have weak support for floating-point operations and high power consumption, the model parameters need to be converted from 32-bit floating-point numbers to low-width fixed-point numbers.

[0090] Specifically, a quantization-aware training tool is introduced into the PyTorch (an open-source deep learning framework) training framework. The weights of convolutional and fully connected layers are quantized into 8-bit integers (INT8), and the membrane potentials and threshold voltages of spiking neurons are quantized into 8-bit or 16-bit fixed-point numbers. Online fine-tuning is then used to adapt the model to the accuracy loss caused by quantization, ensuring that the increase in MAE (Mean Absolute Error) after quantization does not exceed 5%.

[0091] Step S12: HLS (High-Level Synthesis) code conversion. Using Vivado HLS (a high-level synthesis tool), the trained PyTorch model is converted into C++ (Extended C language, a high-level general-purpose programming language) or Verilog (a hardware description language).

[0092] The streaming interface is defined as follows: In the HLS code, the `#pragma HLS INTERFACE axis` directive defines the model's input as an AXI-Stream (a lightweight interface standard in the protocol family specifically designed for high-speed, continuous data streaming) streaming interface. In the field-programmable gate array (FPGA) module logic connection, this interface is configured to interface with the output stream of the preceding Fast Fourier Transform (FFT), thereby receiving the transformed depth-domain A-scan data points.

[0093] To achieve pipeline optimization, the `#pragma HLSPIPELINE II=1` instruction is inserted into the loop bodies of convolutional and fully connected layers. This instruction forces the field-programmable gate array (FPGA) synthesizer to generate pipelined circuitry, ensuring that a new depth data point is processed each clock cycle, thereby achieving nanosecond-level throughput. For fully connected layers, the `#pragma HLS UNROLL` instruction is used to expand matrix multiplications, utilizing the parallel DSP48E (Digital Signal Processing Slice) units within the FPGA to simultaneously calculate the synaptic currents of multiple neurons.

[0094] Step S13: Logic Synthesis and Bitstream Generation. Import the IP (Intellectual Property Core) exported from HLS into the Vivado project. Set the system clock frequency and perform timing analysis to ensure that the critical path delay is less than the clock cycle. Map the logic netlist to specific field-programmable gate array (FPGA) module resources, where BlockRAM (Block Random Access Memory) is used to store weights, and the DSP is used for multiply-accumulate operations. Finally, compile to generate a binary bitstream file and burn it into the FPGA module chip via the JTAG (Joint Test Action Group) interface.

[0095] Through the above deployment method, the streaming convolutional spiking neural network model is transformed into a pure hardware logic circuit inside the field-programmable gate array module, completely freeing it from the constraints of the operating system and CPU instruction set, and realizing hardware-level AI (Artificial Intelligence) inference.

[0096] Figure 6 The flowchart of the real-time feedback control method for optical coherence tomography based on neuromorphic computing provided in this embodiment is shown.

[0097] like Figure 6 As shown, this embodiment provides a real-time feedback control method for optical coherence tomography based on neuromorphic computing. This method relies on the biomimetic "optical reflection arc" hardware architecture described in the previous embodiment and is executed in a pipelined parallel manner within a field-programmable gate array (FPGA) module chip. This method abandons the traditional B-scan image reconstruction process and directly performs streaming processing on the original one-dimensional A-scan signal, specifically including the following steps: Step S21: Acquire and convert the original analog voltage signal from the OCT detector through a hardware direct connection. The field-programmable gate array module directly receives the digital signal stream from the high-speed ADC module via a high-speed serial interface. That is, the system receives the time-encoded original interference signal output by the OCT detector in a point-by-point manner.

[0098] Specifically, this process is completely independent of the operating system scheduling of a general-purpose computer. Data is not cached in DDR memory but is directly written to the FIFO buffer on the field-programmable gate array (FPGA) module. Subsequently, the hardware logic inside the FPGA module performs streaming k-clock resampling or linear interpolation on the data in the FIFO, and then pipelines it through FFT (Fast Fourier Transform) processing to convert the time / wavelength domain interference signal into an A-scan depth signal in the spatial domain, thereby ensuring nanosecond-level deterministic latency for data acquisition and domain transformation.

[0099] Step S22: Surface localization based on A-scan signal to determine the tissue surface location. The preprocessing logic unit inside the field-programmable gate array (FPGA) module reads the A-scan depth signal and performs real-time signal conditioning. First, signal alignment is performed. Using gradient detection or threshold detection algorithms, the position where the signal intensity first significantly increases in the A-scan sequence is identified and marked as the tissue surface index. Then, dynamic alignment is performed. Based on this tissue surface index, the logic unit uses a shift register to perform a translation operation on the subsequent data stream, forcibly aligning the sampling points corresponding to the tissue surface to the logic zero point. This eliminates axial motion artifacts caused by patient breathing or eye tremors, ensuring that the data input to the neural network has a unified spatial reference.

[0100] Step S23: The signal is input point by point into the pre-trained streaming convolutional spiking neural network model. The signal data aligned in step S22 is input point by point into the pre-trained streaming convolutional spiking neural network IP core according to the sampling clock beat.

[0101] Specifically, the model adopts a "wide convolutional preprocessing + multi-layer recursive spiking" architecture. As data points flow in, the convolutional layers inside the model extract local texture features in real time, and the PLIF (Parametric Leaky Integrate-and-Fire) layer updates the membrane potential state in real time using recursive connections, so that inference calculation can begin without waiting for the entire A-scan acquisition to finish.

[0102] Step S24: The model outputs the target depth position index in real time and calculates the remaining thickness. First, the target layer is located. When the streaming convolutional spiking neural network model detects a match between the input data features and the pre-trained target layer features, its output layer voltage will exceed a preset activation threshold. The field-programmable gate array (FPGA) module logic unit captures this trigger moment, records the current global clock count or depth index, and marks it as the target layer depth position index. Then, the subtractor within the logic unit calculates the difference between the target layer index and the surface index in real time and multiplies it by the system's axial physical resolution to obtain the current remaining tissue thickness.

[0103] Step S25: Rule-based threshold judgment and laser truncation control. The remaining tissue thickness calculated in step S24 is compared with a preset safety threshold. To prevent misjudgment caused by single-point noise, a multi-frame sliding window verification mechanism is introduced in this step. The stop condition is only confirmed when the remaining thickness calculated for N consecutive frames is less than the safety threshold. Once the stop condition is confirmed, the field-programmable gate array module immediately controls the GPIO pin level to toggle. This level signal directly acts on the laser's hardware gating interface, physically cutting off the laser energy emission within microseconds, thereby achieving precise "cut-and-stop" control and effectively avoiding the risk of overcutting.

[0104] Next, referring to the accompanying drawings, a real-time feedback control method for optical coherence tomography proposed according to an embodiment of this application is described.

[0105] Figure 7 This is a flowchart illustrating a real-time feedback control method for optical coherence tomography based on neuromorphic computing according to an embodiment of this application. The method is applied to the processing component of the aforementioned real-time feedback control system for optical coherence tomography based on neuromorphic computing.

[0106] like Figure 7 As shown, the real-time feedback control method for optical coherence tomography includes the following steps: In step S201, the analog voltage signal corresponding to the original optical interference signal of the biological tissue is acquired.

[0107] It is understood that the analog voltage signal received by the acquisition module in this application embodiment is used for subsequent circuit acquisition and digital processing.

[0108] In step S202, the analog voltage signal is converted into a digital signal stream, the digital signal stream is input into a neural network model, the neural network model outputs a target depth position index, and a control signal is generated based on the target depth position index. The neural network model obtains the target depth position index through point-by-point streaming inference of the digital signal stream.

[0109] It is understood that the embodiments of this application digitize the analog voltage signal into a continuous digital signal stream in real time through analog-to-digital conversion, and then input it into the neural network model for point-by-point streaming inference (i.e., calculation is performed immediately after each data point is received instead of waiting for the entire frame of data), directly outputting the target depth position index representing key features, and generating the corresponding control signal in real time based on the index, thereby realizing low-latency closed-loop processing from signal acquisition to decision control.

[0110] In this embodiment of the application, before inputting the digital signal stream into the neural network model, the method further includes: performing resampling or linear interpolation on the digital signal stream to correct the nonlinearity of the wavenumber space; performing a fast Fourier transform on the corrected digital signal stream to convert the digital signal stream into an amplitude scanning signal stream; performing real-time surface localization on the amplitude scanning signal stream to obtain the surface position, determining the surface depth position index of the biological tissue, and aligning the amplitude scanning signal stream based on the surface position; and inputting the aligned amplitude scanning signal stream point by point into the neural network model.

[0111] In spectral domain OCT, the pixels of the spectrometer detector are uniformly distributed along the wavelength but nonlinear in the wavenumber space. Fourier transform requires spatially uniform data distribution; otherwise, it leads to depth blurring and reduced resolution. Therefore, resampling or linear interpolation is needed to remap the non-uniformly sampled wavelength data to uniformly distributed wavenumber data to ensure subsequent imaging clarity. The amplitude scan signal stream is a one-dimensional data sequence representing the reflectivity distribution of biological tissue along the beam propagation direction (depth axis) after Fast Fourier Transform processing. Each data point represents the reflectivity intensity of the tissue at a specific depth. The entire curve constitutes a depth profile of the tissue; real-time surface localization is the rapid identification of the outermost layer of biological tissue in the amplitude scan signal stream using algorithms; alignment is the process of shifting the amplitude scan signal stream based on the biological tissue surface depth position index obtained from real-time surface localization, forcibly fixing the tissue surface at the same depth index position in the data stream. The purpose is to eliminate the overall displacement caused by probe jitter or tissue movement, ensuring that the data input to the neural network has a unified coordinate system, so that the network only needs to focus on the relative structural features inside the tissue, rather than the absolute position, thereby greatly improving the inference accuracy and robustness of the model.

[0112] It is understood that, in the embodiments of this application, before the digital signal stream enters the neural network model, the wavenumber space nonlinearity of the digital signal stream needs to be corrected by resampling and linear interpolation to ensure the accuracy of the spectrum analysis. Then, the frequency domain interference stream is converted into an amplitude scan signal stream (A-scan) characterizing the tissue depth reflection distribution using fast Fourier transform. Next, each frame of A-scan is used for real-time surface localization to capture the position of the biological tissue surface, determine the biological tissue surface depth position index, and perform depth alignment of the signal accordingly to eliminate motion artifacts. Finally, the normalized and aligned amplitude signal is streamed point by point into the neural network model to achieve high-precision depth feature inference.

[0113] In this embodiment, the neural network model outputs a target depth location index, including: using a one-dimensional convolutional layer to extract features from the input current data point and its neighborhood data, and outputting simulated feature values; using spiking neurons to encode the simulated feature values ​​into pulse signals; using recursive connections to combine the membrane potential state of the previous moment for temporal memory update; using an output decoding layer to map the pulse features into the confidence voltage value corresponding to the current depth in real time, and when the voltage value exceeds the activation threshold, marking the current moment as the target depth location index.

[0114] The activation threshold is a fixed voltage threshold that is preset and set according to actual needs. It is not specifically limited here. When the confidence voltage value output by the decoding layer exceeds the threshold instantaneously, the depth corresponding to the current moment is determined to be the target depth position index.

[0115] It is understood that the neural network model in this application embodiment first uses a one-dimensional convolutional layer to extract the local waveform features of the current data point and its neighborhood and outputs simulated values. Then, it encodes these simulated features into discrete pulse signals through spiking neurons and combines them with the membrane potential state of the previous moment through recursive connections to achieve temporal memory update, thereby capturing the dynamic evolution of the signal. Finally, the pulse sequence is restored in real time to a confidence voltage value representing the probability through the output decoding layer. Once the voltage exceeds the preset activation threshold, the current moment is locked and marked as the target depth position index, thus realizing low-latency and high-precision depth positioning based on event-driven methods.

[0116] In this embodiment, before inputting the digital signal stream into the neural network model, the method further includes: obtaining a training dataset for the neural network model, wherein the training dataset includes training samples and ground truth labels, the training samples are amplitude scan signal streams, and the ground truth labels are position labels corresponding to the amplitude scan signal streams; training the neural network model using the training samples in the training dataset, during the training process, calculating the training loss using the ground truth labels and the predicted target depth position index of the output of the neural network model, and updating the model parameters of the neural network model according to the training loss; during the training process, randomly generating a cut-off depth, removing data of the length corresponding to the cut-off depth at the front end of the amplitude scan signal stream in the training samples, and shifting the remaining data of the amplitude scan signal stream forward to align it, while simultaneously adjusting the position labels of the training samples.

[0117] Among them, training loss is used to quantify the difference or error between the current predicted target depth location index and the ground truth label of the neural network. The larger the training loss value, the less accurate the model is; the smaller the loss value, the more accurate the model is. Model parameter update refers to automatically adjusting the connection weights and biases between neurons inside the neural network based on the calculated training loss, using the backpropagation algorithm and optimizer. In biological tissue imaging, as the surgery progresses, the surface of the biological tissue is removed, causing the original deep structure to become a new surface layer. The overall depth distribution of the signal will change drastically. During the training process, this process is simulated, and a resection depth is randomly determined, which improves the robustness of the model in actual dynamic surgical scenarios, enabling it to adapt to the constantly changing situation of the tissue surface.

[0118] It is understood that, before the formal deployment of the neural network model, this application embodiment needs to construct a training dataset containing amplitude scanning signal streams and their corresponding ground truth location labels, and introduce a data augmentation strategy to simulate the dynamic process of surgery: during training, the resection depth is randomly generated, the data of the corresponding length of the signal front end is truncated and the remaining part is moved forward to align, and the labels are simultaneously corrected to reflect the new relative position, thereby simulating the real scene after the tissue surface is removed; subsequently, the neural network is iteratively trained using these dynamically augmented samples, the training loss is calculated by comparing the predicted depth index output by the model with the ground truth label, and backpropagation is used to update the model parameters accordingly, ultimately giving the model the generalization ability to accurately locate the target depth even in the dynamic environment of the tissue surface.

[0119] In this embodiment of the application, generating a control signal based on a target depth location index includes: receiving a biological tissue surface location index; calculating the difference between the biological tissue surface location index and the target depth location index; calculating the remaining tissue thickness based on the difference and physical resolution; comparing the remaining tissue thickness with a preset safety threshold; obtaining a quantity threshold; and generating a control signal when it is identified that the remaining tissue thickness of multiple consecutive sampling frames is less than the safety threshold and the number of multiple sampling frames is greater than the quantity threshold.

[0120] Physical resolution refers to the actual physical distance represented by each data point (pixel) in the depth direction of an imaging system (such as OCT) (e.g., each pixel represents 5 micrometers). It converts the index difference in the digital signal stream into a length unit with actual physical meaning (such as millimeters or micrometers). Remaining tissue thickness represents the actual tissue layer thickness from the current tissue surface to the target deep structure that needs to be protected or manipulated. This is the core indicator for judging the safety of surgery, and the calculation method is the target depth index. (Surface position index) × physical resolution; The safety threshold is a pre-set critical thickness value. For example, if a safety margin of 20 micrometers is retained, the safety threshold is set to 20 micrometers. The specific setting is based on the needs of the municipal bureau and is not specifically limited here; The quantity threshold is a pre-set integer counter threshold used to specify the minimum number of frames for continuous alarm triggering. Since biological tissues may have slight jitter or signal noise, the anomaly of a single frame of data may be an instantaneous false positive. Only when the state of remaining thickness less than the safety threshold continues to appear in multiple consecutive sampling frames and the cumulative number of frames exceeds the quantity threshold is it confirmed as a real risk.

[0121] It is understood that the embodiments of this application acquire the surface location index of biological tissue in real time, calculate the difference between it and the target depth location index, and convert it into the actual remaining tissue thickness in combination with the physical resolution of the system. Then, the remaining tissue thickness is compared with a preset safety threshold to assess the risk, and a time-dimensional filtering logic is introduced: only when the abnormal state of remaining thickness below the safety threshold is detected to exist continuously in multiple consecutive sampling frames, and the cumulative number of frames exceeds the quantity threshold, is it determined that the stopping condition is met and a control signal is generated, thereby effectively eliminating instantaneous noise interference and ensuring the high safety and reliability of the surgical operation.

[0122] It should be noted that the control signal design in this embodiment has a low-level latching characteristic, which means that once the stop state is triggered due to the detection that the stop requirement is met, the output control signal (low level) will immediately enter the latching mode, forcibly keeping the output control signal as a low-level stop signal, completely shielding any subsequent input signal fluctuations or recovery attempts (even if the monitoring data temporarily shows normal, the laser will not automatically restart); the core function of this mechanism is to build an insurmountable safety barrier, completely eliminating the risk of laser erroneous restart caused by signal noise, transient interference or logic jitter, and ensuring that after the stop state is triggered, the system can only be unlocked and resumed operation after the operator or host computer issues a clear external reset command for manual confirmation and intervention, thereby maximizing the absolute safety of the surgical process.

[0123] In step S203, the energy output of the surgical laser guided by optical coherence tomography is controlled by a control signal.

[0124] It is understood that the embodiments of this application use control signals as core instructions to adjust the therapeutic laser energy output under optical coherence tomography in real time. When the remaining tissue thickness is detected to be lower than the safety threshold and the control signal is triggered, the laser power attenuation or emergency cutting operation will be executed immediately to ensure that the laser only acts on the target layer in precision surgery, effectively preventing tissue perforation and deep damage caused by energy overload or depth misjudgment.

[0125] The real-time feedback control method for optical coherence tomography (OCT) based on neuromorphic computing proposed in this application can acquire analog voltage signals corresponding to the original optical interference signals of biological tissues through an acquisition module; convert the analog voltage signals into digital signal streams through a processing module, input the digital signal streams into a neural network model, output the target depth position index from the neural network model, and generate control signals based on the target depth position index; and use an execution module to control the energy output of the surgical laser guided by OCT according to the control signals, thereby simulating the reflex arc mechanism of the biological nervous system and constructing a biomimetic optical reflex arc architecture. This method features low computational latency, low computational power consumption, and strong anti-interference capabilities, and can meet the real-time requirements of millisecond-level or even microsecond-level laser surgery.

[0126] It should be noted that the foregoing explanation of the embodiment of the real-time feedback control system for optical coherence tomography based on neuromorphic computing also applies to the real-time feedback control method for optical coherence tomography based on neuromorphic computing in this embodiment, and will not be repeated here.

[0127] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0128] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0129] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0130] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0131] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0132] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A real-time feedback control system for optical coherence tomography based on neuromorphic computing, characterized in that, include: An optical coherence tomography (OCT) device and a processing module, wherein the OCT device includes an acquisition module and an execution module, wherein... The acquisition module is used to acquire the analog voltage signal corresponding to the original optical interference signal of biological tissue; The processing module is used to convert the analog voltage signal into a digital signal stream, input the digital signal stream into a neural network model, output a target depth position index from the neural network model, and generate a control signal based on the target depth position index. The neural network model obtains the target depth position index through point-by-point streaming inference of the digital signal stream. An execution module is used to control the energy output of the surgical laser guided by optical coherence tomography according to the control signal.

2. The real-time feedback control system for optical coherence tomography based on neuromorphic computing according to claim 1, characterized in that, The acquisition module includes an optical coherence tomography (OCT) probe and an OCT detector. The OCT probe and the OCT detector are optically linked. The OCT probe acquires the raw optical interference signal of biological tissue, and the OCT detector outputs the analog voltage signal.

3. The real-time feedback control system for optical coherence tomography based on neuromorphic computing according to claim 1, characterized in that, The execution module includes a laser control gate and a laser probe. The laser control gate is an acousto-optic modulator or a seed source trigger interface, which is directly controlled by the control signal output by the processing module.

4. The real-time feedback control system for optical coherence tomography based on neuromorphic computing according to claim 1, characterized in that, The processing module includes an analog-to-digital converter module and a field-programmable gate array (FPGA) module. The analog-to-digital converter module converts the analog voltage signal into a digital signal stream. A hardware direct connection data path is established between the analog-to-digital converter module and the FPGA module. The hardware direct connection data path directly inputs the digital signal stream to the FPGA module in a streaming manner without going through general-purpose computer memory and operating system scheduling.

5. The real-time feedback control system for optical coherence tomography based on neuromorphic computing according to claim 1, characterized in that, The neural network model is set in the field-programmable gate array module. The neural network model adopts a unidirectional streaming architecture, including a wide convolutional preprocessing module, a multi-layer recursive spiking neuron module and an output decoding module connected in sequence. The wide convolutional preprocessing module is configured with a one-dimensional convolutional layer for extracting local texture features of the input signal; The multi-layer recursive spiking neuron module includes at least two spiking neuron layers, and residual connections are provided between the spiking neuron layers. The multi-layer recursive spiking neuron module is configured to directly superimpose the input current of the shallow spiking neuron layer to the input end of the deep spiking neuron layer. The multi-layer recursive spiking neuron module is also configured with recursive connections for transmitting the pulse state at the current moment back to the input end at the next moment. The output decoding module is used to output the target depth position index.

6. A real-time feedback control method for optical coherence tomography based on neuromorphic computing, characterized in that, The method, applied to the processing component of the real-time feedback control system for optical coherence tomography based on neuromorphic computing as described in any one of claims 1-5, includes the following steps: Obtain the analog voltage signal corresponding to the original optical interference signal of biological tissue; The analog voltage signal is converted into a digital signal stream, the digital signal stream is input into a neural network model, the neural network model outputs a target depth position index, and a control signal is generated based on the target depth position index. The neural network model obtains the target depth position index through point-by-point streaming inference of the digital signal stream. The energy output of the surgical laser guided by optical coherence tomography is controlled using the control signal.

7. The real-time feedback control method for optical coherence tomography based on neuromorphic computing according to claim 6, characterized in that, Before inputting the digital signal stream into the neural network model, the method further includes: Resampling or linear interpolation is performed on the digital signal stream to correct for nonlinearity in the wavenumber space; Perform a Fast Fourier Transform on the corrected digital signal stream to convert the digital signal stream into an amplitude scanning signal stream; The amplitude scan signal stream is subjected to real-time surface localization to obtain the surface position, the surface depth position index of the biological tissue is determined, and the amplitude scan signal stream is aligned based on the surface position. The aligned amplitude scanning signal stream is input point by point into the neural network model.

8. The real-time feedback control method for optical coherence tomography based on neuromorphic computing according to claim 7, characterized in that, The neural network model outputs the target depth location index, including: A one-dimensional convolutional layer is used to extract features from the input current data point and its neighborhood data, and output simulated feature values. The simulated feature values ​​are encoded into pulse signals using spiking neurons; The temporal memory is updated by using recursive connections combined with the membrane potential state of the previous time step; The output decoding layer is used to map the pulse features to the confidence voltage value corresponding to the current depth in real time. When the voltage value exceeds the activation threshold, the current time is marked as the target depth position index.

9. The real-time feedback control method for optical coherence tomography based on neuromorphic computing according to claim 6, characterized in that, Before inputting the digital signal stream into the neural network model, the following steps are also included: Obtain a training dataset for a neural network model, wherein the training dataset includes training samples and ground truth labels, the training samples are amplitude scanning signal streams, and the ground truth labels are position labels corresponding to the amplitude scanning signal streams; The neural network model is trained using training samples in the training dataset. During the training process, the training loss is calculated using the ground truth labels and the predicted target depth location index of the output of the neural network model. The model parameters of the neural network model are updated based on the training loss. During training, a cut-off depth is randomly generated, and data of the length corresponding to the cut-off depth at the front end of the amplitude scan signal stream in the training samples is removed. The remaining data of the amplitude scan signal stream is then moved forward to align, and the position labels of the training samples are adjusted synchronously.

10. The real-time feedback control method for optical coherence tomography based on neuromorphic computing according to claim 6, characterized in that, Generate a control signal based on the target depth position index, including: Receives the location index of biological tissue surface; Calculate the difference between the biological tissue surface location index and the target depth location index, and calculate the remaining tissue thickness based on the difference and physical resolution. The remaining tissue thickness is compared with a preset safety threshold; A quantity threshold is obtained. When it is identified that the remaining tissue thickness of multiple consecutive sampling frames is less than the safety threshold, and the number of the multiple sampling frames is greater than the quantity threshold, the control signal is generated.

Images