1. The Clinical & Technical Challenge

• Workflow Phase Recognition: Automatically logging surgical steps to assist OR teams and optimize scheduling.
• Skill Assessment & Ergonomics: Providing objective feedback on tool movement patterns to trainees.
• Intraoperative Safety: Alerting surgical staff to off-camera tools or hazardous instrument proximity.

• Dynamic occlusion by internal organs and tissue
• Severe specular glare from metallic surfaces
• Surgical smoke, blood, and fluid misting
• Rapid camera movement and varying focal lengths

2. Why YOLO? The Latency vs. Accuracy Equation

Architectural Highlights:

1. Anchor-Free Detection: Modern YOLO backbones reduce manual hyperparameter tuning while boosting detection speed for objects with extreme scale variations.
2. Feature Aggregation Networks: Advanced FPN and PAN modules preserve fine-grained spatial features, keeping tools detectable even through heavy surgical smoke or partial occlusion.

3. Key Research Insights & Benchmarks

4. Innovation Takeaways for AI Startups

1. Adopt a Hybrid Edge-Cloud Architecture: Deploy lightweight, high-FPS YOLO models directly on edge hardware in the OR for zero-latency visual overlays, while streaming video to larger, high-mAP backbones in the cloud for post-operative analytics.
2. Focus on Domain-Specific Augmentation: Data augmentation techniques designed specifically for surgical smoke, fluid misting, and glare yield greater real-time accuracy gains than simply expanding model size.
3. Build Modular CV Pipelines: Designing a flexible model abstraction layer allows AI platforms to upgrade to newer open-source YOLO architectures seamlessly without refactoring core software logic.

Strategic Summary

1. Executive Summary

We have developed a highly efficient, end-to-end artificial intelligence pipeline that automates Racing Bib Number (RBN) detection and recognition from dynamic, real-world racing environments.

By replacing expensive, hardware-heavy RFID chip timing systems and slow, post-event manual photo tagging, our proprietary framework achieves state-of-the-art accuracy with an unprecedented end-to-end inference speed of 23.5 milliseconds. Crucially, our innovation circumvents the traditional AI bottleneck: achieving enterprise-grade reliability using a highly curated, data-augmented dataset that requires 99.7% less training data than historical deep learning benchmarks.

2. The Market Challenge

The global endurance sports market (marathons, triathlons, cycling, and track events) relies heavily on tracking athlete participation for timing, live broadcasting, and commercial photo sales. The current industry standard faces two massive friction points:

1. Hardware Dependencies (RFID Chips): Magnetic and RFID timing mats are logistically difficult to deploy, prone to localized failure (e.g., missed reads in crowded finishes), and provide no visual media asset tagging.

2. The “Big Data” AI Bottleneck: Previous attempts to use deep learning for visual bib recognition required astronomical datasets (nearly 500,000 manually labeled images), creating an insurmountable barrier to entry for lightweight, agile commercial deployments.

3. Latency Barriers: Existing computer vision models suffer from heavy processing latencies (ranging from 750 ms to over 2,200 ms per image), rendering them completely non-viable for real-time video stream integration or live broadcast overlays.

3. The Innovation & Technology Stack

Our startup re-engineered the RBN pipeline by splitting it into a specialized, ultra-fast dual-stage architecture optimized for low-latency execution:

[Live Video Stream / Photo Input] 
               │
               ▼
   ┌───────────────────────┐
   │ Real-Time Localization│  <-- Custom-optimized YOLOv4 Object Detector
   └───────────────────────┘
               │
               ▼
   ┌───────────────────────┐
   │ Dynamic Patch Extraction│ <-- High-speed mathematical bounding-box crop
   └───────────────────────┘
               │
               ▼
   ┌───────────────────────┐
   │ Text Extraction Engine│  <-- Deep LSTM-based Tesseract OCR (PSM 13)
   └───────────────────────┘
               │
               ▼
 [Structured Output Data: "Bib #1370"]

Stage 1: Ultra-Fast Localization (YOLOv4)

We utilized a custom-tuned YOLOv4 (You Only Look Once) deep learning model, initiated with pre-trained convolutional weights (yolov4.conv.137), to isolate RBN bounding boxes. To train this model with maximum efficiency, we pioneered advanced data augmentation pipelines, notably combining Cutmix and Mosaic data augmentation. This allowed our network to learn diverse racing context cues (road, cross-country, indoor/outdoor track) from limited imagery.

Stage 2: Context-Aware Text Extraction (Tesseract OCR + LSTM)

Once localized, the image patch is dynamically cropped and passed to an optimized Tesseract 4 OCR engine driven by Long Short-Term Memory (LSTM) neural networks. We optimized this stage by enforcing a strict numerical whitelist (0-9) and deploying Page Segmentation Mode (psm = 13), forcing the engine to read the crop as a raw, single text line.

4. Key Performance Indicators & Results

Our proprietary model underwent rigorous testing against both our native multi-terrain sports datasets and standard industry benchmarks (RBNR dataset), demonstrating flawless technical performance:

Technical Accuracy Metrics

• Mean Average Precision (mAP): 0.935 — indicating exceptional spatial accuracy in locating bibs.

• Recall / Precision / F1-Score: 0.91 / 0.88 / 0.89 — showcasing highly balanced detection that limits false positives.

• Full Digit Text Recognition Rate: 71% exact full-string match on detected bibs, outperforming previous academic baselines on raw text extraction.

Computational Efficiency & Speed

• Detection Phase (GPU): 21 ms

• Recognition Phase (CPU/LSTM): 2.5 ms

• Total End-to-End Latency: 23.5 ms

5. Competitive Advantage & Commercial Disruption

Our startup’s technology introduces three distinct, unfair competitive advantages to the B2B sports tech market:

┌───────────────────────────────────────────────────────────────────────────┐
│                          COMPETITIVE LANDSCAPE                            │
├──────────────────────┬─────────────────────────────┬──────────────────────┤
│ Metric               │ Legacy AI Frameworks        │ Our Startup's Model  │
├──────────────────────┼─────────────────────────────┼──────────────────────┤
│ Required Training    │ 498,385 Labeled Images      │ 1,374 Labeled Images │
│ Data Scale           │                             │ (99.7% reduction!)   │
├──────────────────────┼─────────────────────────────┼──────────────────────┤
│ Processing Latency   │ 750 ms – 2,240 ms           │ 23.5 ms              │
│                      │ (Batch process only)        │ (True Real-Time)     │
├──────────────────────┼─────────────────────────────┼──────────────────────┤
│ Operational Viability│ Post-Event Processing Only  │ Live Stream & Edge   │
│                      │                             │ Deployable           │
└──────────────────────┴─────────────────────────────┴──────────────────────┘

1. Unrivaled Data Efficiency (The “Small Data” Edge): While competitors require expensive, half-million-image datasets, our data augmentation strategy achieves comparable or superior accuracy using just 1,374 images. This slashes our R&D costs and allows us to rapidly adapt the model to new client requirements (e.g., cycling plates or obstacle course mud-splattered bibs) with minimal training overhead.

2. True Real-Time Capability: At 23.5 ms per frame, our framework operates comfortably within the standard 30–60 FPS video broadcast window. This transforms bib recognition from a slow, post-race batch task into a live tool capable of driving immediate broadcast graphics, digital race tracking apps, or instant security flagging.

3. Hardware-Agnostic, Edge-Ready Footprint: By decoupling the architecture—keeping the detection on lightweight GPU nodes and offloading text recognition to 2.5 ms CPU threads—our software can be deployed on cost-effective edge-computing devices at the finish line, dropping cloud infrastructure costs dramatically.

6. Future Roadmap & Scale

Moving forward, our startup is expanding the commercial capabilities of this architecture into:

• Multi-Modal Athlete Re-Identification (Re-ID): Fusing the RBN detector with secondary physical attributes (jersey colors, posture, and facial geometry) to maintain tracking even when an athlete’s bib is temporarily obscured by arms, weather, or other runners.

• Automated Media Monetization Pipelines: Instant, real-time watermarking and delivery of personalized race photos directly to runners’ smartphones seconds after they cross intermediate milestones.

Technology Domain: Computer Vision (CV), Deep Learning, Edge Computing, Optical Character Recognition (OCR)

Target Market: Endurance Sports Event Management, Broadcast Media, Sports Photography, and Athletic Analytics

Read more about work here.

At OptimumT, we’re on a mission to push the boundaries of AI, especially where it matters most – saving lives. We’ve all seen the headlines about groundbreaking AI models, boasting incredible performance in labs and on massive public datasets. But what happens when these “State-of-the-Art” models face the gritty, unpredictable reality of critical applications like Computer-Aided Laparoscopy (CAL)?

A recent study, “Performance Analysis of YOLO-NAS SOTA Models on CAL Tool Detection”, dives deep into this very question, and the results are a

powerful wake-up call for the entire AI community!

The Hype vs. Reality Check for YOLO-NAS

You’ve probably heard of YOLO-NAS (You Only Look Once – Neural Architecture Search), touted by its creators as a leap forward in object detection, promising superior accuracy and computational efficiency. The theory is sound: using advanced Neural Architecture Search (NAS) to automate the design of optimal AI models should yield incredible results.

However, when our researchers put these models to the test on a real-world CAL dataset – detecting surgical tools in intricate laparoscopic procedures – the narrative took an unexpected turn.

The Shocking Truth:

Despite the impressive claims and initial benchmarks on general datasets like COCO, the YOLO-NAS models (small, medium, and large variants) performed inferiorly compared to established State-of-the-Art YOLO models such as YOLOv7 and YOLOv8n on the CAL tool detection task. In fact, their performance was described as “dismal” in certain areas, particularly in detecting the “bipolar” tool.

This isn’t just about numbers on a chart; it’s about the critical difference between theoretical performance and reliable operation in high-stakes environments. CAL, with its challenges like smoke, blood, reflections, and complex backgrounds, demands incredibly robust and accurate AI systems.

What Does This Mean for the Future of AI in Healthcare?

This research underscores a vital principle that guides our work at [Your Startup Name]: Rigorous, real-world validation is paramount. It’s not enough for an AI model to perform well on generalized datasets; it must prove its mettle in the specific, complex environments where it will be deployed.

At OptimumT, we are committed to:

• Deep Domain Expertise: Understanding the nuances of critical applications, such as surgical assistance, is at our core.
• Battle-Tested AI: Our models are not just trained; they are hardened through extensive testing on specialized, real-world datasets, ensuring they perform when it truly counts.
• Ethical Deployment: We believe in transparency and robust evaluation to build trust and ensure the safety and effectiveness of AI in sensitive fields.

This study is a powerful reminder that while AI is evolving at an exhilarating pace, the true measure of its impact lies in its ability to deliver reliable and superior performance in the face of real-world complexity.

Join us as we build the next generation of AI that doesn’t just look good on paper, but performs brilliantly when lives are on the line!

#AI #ArtificialIntelligence #HealthcareAI #ObjectDetection #YOLONAS #MedicalTechnology #Startup #Innovation #RealWorldAI #CAL #DeepLearning #ComputerVision

At OptimumT, we are committed to harnessing the power of artificial intelligence to transform the healthcare landscape. Our latest research, titled “Cholecystectomy Surgical Instrument Detection Using Variants of YOLOv8,” marks a significant milestone in our journey to enhance surgical precision and efficiency through advanced technology. The article was recently published in the recent proceedings of the prestigious Irish Signals and Systems Conference (ISSC), 2024. Here is another link to our work.

Here is the corresponding oral presentation of this work.

The Challenge in Modern Surgery

Surgery is an intricate and demanding field that requires not only exceptional skill but also the ability to make quick, accurate decisions in high-pressure environments. As the number of surgical procedures continues to rise globally, the need for innovative solutions that can assist surgeons in real-time has never been more critical. This is where object detection technologies come into play, offering the potential to revolutionize how surgical instruments are tracked and utilized during operations.

Enter YOLOv8: The Next Generation of Object Detection

The You Only Look Once (YOLO) algorithm has long been a leader in the realm of object detection, and its latest iteration, YOLOv8, brings forth groundbreaking advancements. Our research focuses on leveraging the capabilities of YOLOv8 to improve the detection of surgical instruments during cholecystectomy procedures, a common yet complex surgery.

Key Features of YOLOv8:
Superior Detection Accuracy: YOLOv8 is designed to enhance the accuracy of instrument detection, ensuring that surgical tools are identified correctly even in challenging visual conditions.
Rapid Inference Speed: The algorithm’s efficiency allows for real-time processing of laparoscopic video feeds, providing surgeons with immediate feedback and insights.
Robust Performance Across Variants: Our study explores the various models within the YOLOv8 family, demonstrating their effectiveness in diverse surgical scenarios.

Research Insights: A New Benchmark in Surgical Instrument Detection

In our study, we utilized the m2cai16-tool-locations dataset, which includes 2,811 frames from 10 videos, annotated with 3,141 instances of seven different surgical instruments. By training multiple YOLOv8 models on this dataset, we achieved remarkable results that not only highlight the algorithm’s capabilities but also set a new standard for surgical instrument detection.

Benefits of Our Findings:
Enhanced Surgical Workflow: The integration of YOLOv8 into CAL systems allows for automated tool recognition, reducing the cognitive load on surgeons and enabling them to focus on the procedure.
Improved Patient Safety: Accurate instrument tracking minimizes the risk of surgical errors, ultimately enhancing patient safety and outcomes.
Contribution to the Surgical Community: Our research not only benefits surgeons but also contributes to the ongoing development of the YOLO algorithm, paving the way for future advancements in object detection.

Looking Ahead: The Future of AI in Healthcare

As we continue to explore the potential of AI in healthcare, the implications of our findings are profound. The integration of YOLOv8 in surgical instrument detection represents a significant leap forward, not just for surgeons but for the entire healthcare ecosystem. By embracing these advancements, we can redefine the standards of surgical precision and improve patient care.

At OptimumT, we are excited about the future of AI in healthcare and are committed to driving innovation that enhances surgical practices. Join us on this journey as we continue to push the boundaries of technology and work towards a future where surgical excellence is within reach for all.

Stay tuned for more updates on our research and developments as we strive to make a meaningful impact in the world of healthcare through artificial intelligence.

In the realm of medical technology, the intersection of artificial intelligence and surgical robotics is paving the way for groundbreaking advancements. At the forefront of this revolution is the innovative research presented in the paper “NeuroEvolution of Capsule Networks for Computer-Aided Laparoscopy.” This cutting-edge study explores how neuroevolutionary techniques can enhance the capabilities of capsule networks, ultimately transforming the landscape of laparoscopic surgery.

The Challenge: A Growing Need for Surgical Precision

Every year, approximately 330 million surgeries are performed globally, yet a staggering backlog of 150 million surgeries remains pending. As the demand for surgical procedures continues to rise, the need for precision and efficiency in the operating room has never been more critical. Traditional surgical methods often fall short in providing the necessary accuracy, leading to longer recovery times and increased risks for patients. This is where the integration of advanced technologies, such as computer vision and artificial intelligence, comes into play.

Enter NeuroEvolution: A Game-Changer for Surgical Robotics

The research highlights the application of neuroevolution—a technique that optimizes artificial neural networks (ANNs) using evolutionary algorithms—to develop capsule networks specifically designed for computer-aided laparoscopy (CAL). Unlike conventional convolutional neural networks (CNNs), capsule networks offer a more robust approach to understanding the spatial relationships between objects, making them ideal for surgical tool classification and real-time decision-making during procedures.

By evolving the architecture of these networks, the study demonstrates how surgical robots can achieve unprecedented levels of accuracy in identifying tools and tracking their movements. This advancement not only enhances the surgeon’s ability to perform delicate operations but also significantly reduces the risk of errors, ultimately improving patient outcomes.

The Power of Computer Vision in Surgery

Computer vision is revolutionizing the way surgeries are performed. With the ability to track surgical instruments, detect organs, and analyze the surgical phase, surgeons can operate with greater confidence and precision. The research emphasizes that these advancements enable remote robotic surgery, allowing surgeons to perform procedures from thousands of miles away. This capability opens up new possibilities for healthcare access, particularly in underserved regions where specialized surgical expertise may be lacking.

 Promising Results: A Glimpse into the Future

The findings of this research are not just theoretical; they represent a tangible step toward the future of surgery. By applying neuroevolution to capsule networks, the study showcases promising results in tool classification for CAL, indicating that the full potential of this technology is yet to be realized. As we continue to explore the capabilities of AI in healthcare, the implications for surgical practice are profound.

Join the Revolution: Embrace the Future of Surgery

At OptimumT, we are excited about the potential of applied artificial intelligence in transforming the medical field. The insights gained from the “NeuroEvolution of Capsule Networks for Computer-Aided Laparoscopy” paper inspire us to push the boundaries of innovation and explore new ways to enhance surgical practices. As we stand on the brink of a new era in healthcare, we invite you to join us in embracing the future of surgery.

Conclusion: A New Dawn for Surgical Robotics

The integration of neuroevolution and capsule networks into CAL represents a thrilling advancement in surgical technology. As we continue to harness the power of artificial intelligence, the possibilities for improving patient care and surgical outcomes are limitless. Together, we can revolutionize the way surgeries are performed, making them safer, more efficient, and accessible to all.

In recent years, the healthcare industry has witnessed a remarkable transformation driven by technological advancements. Among these, artificial intelligence (AI) has emerged as a game-changer, offering innovative solutions across various medical fields. One area where AI’s impact is particularly profound is in cardiorespiratory rehabilitation. As someone who has extensively reviewed the applications of AI in this domain, I’m excited to share the promising developments and future potential that AI brings to cardiorespiratory rehabilitation.

 The Growing Need for Advanced Rehabilitation Solutions

Cardiorespiratory conditions, including chronic obstructive pulmonary disease (COPD) and heart failure, pose significant challenges to patients and healthcare providers alike. Traditional rehabilitation programs, while effective, often face limitations such as accessibility, personalization, and continuous monitoring. This is where AI steps in, offering solutions that enhance the efficiency and effectiveness of rehabilitation programs.

AI-Powered Personalization and Monitoring

One of the standout benefits of AI in cardiorespiratory rehabilitation is its ability to provide personalized care. AI algorithms can analyze vast amounts of patient data, including medical history, current health status, and response to treatment, to create tailored rehabilitation plans. This level of personalization ensures that each patient receives the most appropriate and effective therapy, improving their chances of recovery.

Moreover, AI-powered wearable devices and sensors enable continuous monitoring of patients’ vital signs and physical activity. These devices collect real-time data, allowing healthcare providers to track progress and make necessary adjustments to the rehabilitation plan. The integration of AI ensures that even subtle changes in a patient’s condition are detected early, enabling timely interventions and preventing potential complications.

Enhancing Patient Engagement and Adherence

Patient engagement and adherence to rehabilitation programs are crucial for successful outcomes. AI-driven applications, such as virtual coaches and gamified exercise routines, make rehabilitation more engaging and enjoyable. These tools provide patients with real-time feedback, encouragement, and motivation, making it easier for them to stick to their rehabilitation regimen.

Virtual reality (VR) and augmented reality (AR) technologies, powered by AI, are also making their way into cardiorespiratory rehabilitation. These immersive experiences not only make exercises more enjoyable but also help patients practice breathing techniques and physical activities in a controlled, simulated environment.

Predictive Analytics and Outcome Prediction

Another significant advantage of AI in cardiorespiratory rehabilitation is its ability to predict outcomes. By analyzing historical data and identifying patterns, AI can predict how a patient is likely to respond to a specific rehabilitation program. This predictive capability allows healthcare providers to proactively address potential issues, optimizing the rehabilitation process for better results.

Furthermore, AI-driven predictive analytics can assist in identifying patients at higher risk of adverse events, enabling preemptive measures to mitigate those risks. This level of foresight is invaluable in ensuring patient safety and enhancing the overall effectiveness of rehabilitation programs.

The Future of Cardiorespiratory Rehabilitation

The integration of AI in cardiorespiratory rehabilitation is still in its early stages, but the potential is immense. As AI technologies continue to evolve, we can expect even more sophisticated applications that will further revolutionize the field. From advanced predictive models to fully autonomous rehabilitation programs, the future of cardiorespiratory rehabilitation looks promising.

In conclusion, the applications of AI in cardiorespiratory rehabilitation are transforming the way we approach patient care. By offering personalized, engaging, and predictive solutions, AI is enhancing the effectiveness of rehabilitation programs and improving the quality of life for patients with cardiorespiratory conditions. As we continue to explore and harness the power of AI, we move closer to a future where optimal recovery is within reach for everyone.

If you’re interested in learning more about the specifics of these AI applications and the latest research in the field, we encourage you to read our full review paper. Together, let’s embrace the future of healthcare with AI-powered rehabilitation solutions.

One of our recent projects was to apply novel computer vision techniques to come up with a scheme to detect tool wear during machining. Machining is an industrial manufacturing process that is applied to shape materials. Specialized materials are used as tools for machining. During the process, it is inevitable that the tools shall undergo wear and tear. Heat and abrasion are the usual causes of degradation.  Detecting tool wear in real-time is quite important for the process engineers.

In our approach we employed convolutional neural networks to train a model to predict wear and tear of machining tools. Our work was published in a very important journal related to manufacturing. The article can be found as follows.

Detection of tool wear during machining by designing a novel 12-way 2-shot learning model by applying L2-regularization and image augmentation – The International Journal of Advanced Manufacturing Technology

Tool wear monitoring is regarded as an incredibly important aspect of improving the surface integrity of machined components in the manufacturing sector. This research study performed operations using twelve different types of drilling and milling tools. The worn tools ranging from grade-1 to grade-5 were categorized based on tool wear severity by measuring the flank wear land width of each tool.

Imagine a world where your CNC machines warn you of a dying tool before it wrecks your expensive workpiece, grinds production to a halt, or sparks a safety incident. At OptimumT, we’re turning that vision into reality — today.

🔥 Meet Our Game-Changing 12-Way 2-Shot Learning Model 🔥

🔍 What It Does

• Instantly classifies tool health from “brand-new” all the way to “critically worn.”

• Works for both milling and drilling on high-performance alloys like Inconel 718.

• Tells you exactly when to replace or re-sharpen—saving hours of downtime and thousands in scrap costs.

🎯 Why It’s Revolutionary

• Only TWO images per wear-class needed. No massive data lakes.

• Smarter image augmentation jumps from 24 originals to 240 training samples—all without hefty manual labeling.

• Built on ResNet-18 to “see” force-signal spectrograms and scalograms like a seasoned machinist.

• L2 regularization ensures rock-solid accuracy (87.83% in rigorous tests) and zero overfitting.

🔧 How It Works in a Nutshell

1. 💥 Force Sensing: A dynamometer records real-time force signals as tools cut through tough metals.

2. 🖼️ Visual Transformation: We convert those raw signals into vibrant spectrograms and scalograms—think “sonar maps” of tool wear.

3. 🧠 Meta-Learning Magic: Our few-shot network compares a brand-new image against just two examples per wear level—then instantly knows if a tool is healthy or on its last breath.

4. 🚀 Actionable Insights: Alerts fire off to your CNC control system or maintenance dashboard, so you can swap that milling cutter before it gouges your part.

💪 Why Manufacturers Are Raving

• Zero retraining headaches. Deploy it on a forgotten production line; it learns on the spot.

• Ultra-low data requirements. Securely monitor every spindle without building giant data lakes.

• Real-time, shop-floor intelligence. No more guesswork—your team fixes only what needs fixing, when it needs fixing.

🔮 The Bottom Line
Our 12-way 2-shot learning model isn’t just another proof-of-concept. It’s a factory-floor game-changer that:

• Slashes unplanned downtime

• Drastically reduces scrap & rework

• Boosts throughput and quality

• Empowers maintenance teams with actionable alerts

Ready to supercharge your smart factory? Let’s make sure every tool in your arsenal works at peak performance—every single cut. 💥

👇 Reach out to see a live demo or talk integration—because in the world of precision machining, perfect timing is everything.

#SmartManufacturing #Industry40 #AI #PredictiveMaintenance #FewShotLearning #CNCInnovation #ToolWearDetection #DeepLearning #ManufacturingTech #StartupEnergy

Recently we pushed a new version of GELAB to its GitHub repository. The new version was compiled using Matlab R2023b. The mapper, that is already written in Java, was also updated with a fresh compile. Grammatical evolution is already a well-established machine learning and artificial intelligence technique. The algorithm has been applied to a myriad of real-world problems. It has plenty of applications. Ranging from robotics, computer vision, predictive analytics, game playing, etc., it has already been applied in a lot of niche domains successfully. Having it in Matlab is additional convenience, as it is easy to use. Below is a link to the GitHub repository of GELAB. We hope that you will enjoy working with this software.