#image classification

tl;dr there’s a vaguely defensible reason clone trooper armour isn’t camo, and it’s because droid vision & pattern recognition is so different from humans 

For reasons, I was messing around with machine learning and image processing, and it got me thinking about camouflage. It only works on and for us because of how our human biological vision works, a series of hacks that kept our upstart savannah ape ancestors alive under the light of a GV star.

Droids, though? No evolutionary pressure and no tradeoffs. No grayscale at low light, no limitations in their visible spectrum. They could go all the way up infrared and ultraviolet without crashing their visual input. Their image processing and pattern recognition would be completely different. Camouflage wouldn’t work on droids, not consistently: once they’d figured out the signature for one pattern and shared it around, it’d never work again. Depending on type of camo, it might not work on a given species under a different solar spectrum, either. 

Since the Kaminoans don’t know where and who the clone troopers will be fighting, it’s a better investment for them to focus on armour development rather than camouflage patterns. Camo is too relative to the environment to usually be worth their R&D budget. Maybe they tried for ultraviolet adversarial preturbation patterns instead? Something invisible to the troopers themselves but catastrophic for machine image classification.

We do, however, see some limited use of camo patterned armour in TCW. One is during Landing at Point Rain, where arguably most of the threat is from the aerial Geonosians. It’s also a major offensive with multiple Jedi generals dirtside; maybe it was big enough to be worth changing up the patterns, especially since their major adversaries weren’t droids. 

Source: http://ai.googleblog.com/2019/05/efficientnet-improving-accuracy-and.html

Adversarial Machines

There’s been some interesting developments recently in adversarial training, but I thought it would probably be a good idea to first talk about what adversarial images are in the first place. This Medium article by @samim is an accessible explanation of what’s going on. It references this talk by Ian Goodfellow, asking if statistical models understand the world.

Machine learning can do amazing magical things, but the computer isn’t looking at things the same way that we do. One way to exploit that is by adding patterns that we can’t detect but that create enough of a difference in the data to completely fool the computer. Is it a dog or an ostrich?

There’s been quite a lot of research into finding ways round this problem as well as exploiting it to avoid facial recognition or other surveillance. And, like I said, there’s been some interesting recent developments that I hope to talk about here.

Universal Adversarial Perturbations

Now for the recent developments in adversarial image mis-classifications: fixed patterns that can cause most images to be mis-categorized.

This research by Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, Omar Fawzi, and Pascal Frossard has lead to the discovery that there are patterns that can perturb nearly every image enough to cause an image recognition neural net to fail to identify it most of the time.

Earlier research into adversarial patterns required that you solve an optimization problem for each image. The universal patterns don’t need to do this: they can just add the pattern to every image, which saves a lot of time and effort.

Interestingly, despite the universal patterns being slightly different between different neural net architectures, they are still universal enough that one pattern works pretty well on all of the neural nets.

So if you want to fool, say, a facial recognition system, but you don’t know what neural net it uses, you can just deploy one of these patterns and expect that you’ll likely trick it.

It also tells us more about what is going on under the hood of these neural networks. Understanding why these universal patterns work can help us build better neural networks. The researchers think that the explanation might be because there’s a correlation between different regions of the decision boundary, so the geometry of the decision boundary is what they plan to research in the future.

Infosearch provides image classification services for various industries including healthcare.

In the recent data-saturated healthcare industry, effective quality healthcare requires the capability to interpret and analyze medical images precisely, and at fast rates of transformation, literally saving life and death. Artificial intelligence (AI)-based image classification is used in the medical field using X-rays and MRIs, as well as pathology slides and retinal scans.

But developing and sustaining complex AI models internally is a significant problem to majority of healthcare organizations. This is why more and more hospitals and the research labs, as well as healthcare tech startups, are outsourcing image classification services, and enjoying the rewards of quicker diagnoses, less burden, and better patient results.

This blog post aims to discuss how outsourced AI image classification is becoming an essential element of contemporary healthcare diagnostics, as well as its advantages, potential risks and good practices in its implementation.

What is the healthcare image classification?

An example of a computer vision task is image classification, where the AI gets trained to examine visual data (usually medical data) and place them into a set of previously assigned labels: e.g. it might be asked to categorize the image into one or more of the following categories: malignant tumor, fracture, no abnormality detected.

Medical applications are:

•             Radiology: Categorizing CT, MRI or X-rays scans in search of disease indicators

•             Pathology: detection of cancerous cells in biopsy lesions

•             Ophthalmology: Finding signs of diabetic retinopathy in eye images

•             Dermatology: The characterization of skin wart or moles

•             Cardiology: Diagnosis of echocardiograms in heart conditions

Why Medical Image Classification Outsource?

1. 🕐 Quicker Diagnostics using Scalable Infrastructure

Lower-cost outsourced services can access AI models trained on imaging data and GPU resources, which enable quick processing of voluminous imaging data, required at the time of an emergency or when a clinic is overwhelmed.

2. Lower Costs of Operation and Development

Building, supporting and training AI models internally is expensive as it would involve teams of data scientists, radiologists, and engineering personnel. This expertise is only availed through outsourcing at reduced cost.

3. When combined with the human-machine collaboration, accuracy increased.

Leading outsourced providers commonly utilize both automation of classification as well as human review with a human-in-the-loop approach, thereby increasing performance and diagnostic precision particularly at edge cases or rare disease.

4. Lifelong Education and Model edition

The partnership with outsourcing will allow continuously training the AI models relying on the real-world feedback meaning that the system will gradually improve and not put an extra burden on the internal teams.

Healthcare Diagnostics

🩻 Radiology

AI tags X-rays, MRIs and CT scans to identify:

•             Pneumonia

•             Lung nodules

•             Bleeding in the brain hemorrhages

•             Spinae deformities

Advantage: Triaging of important cases with radiologists.

🔬 Pathology

Image categorisation:

•             Biopsied-cancer cells

•             Blood smear infections

•             Pathologies of a tissue slide

Advantage: Quicker and more effective pathology reviews, in particularly resource-poor regions.

👁️ Ophthalmology

AI is able to recognize:

•             Diabetic retinopathy

•             Glaucoma

•             The age-related macular degeneration (AMD)

Advantage: Early identification and treatment especially in underdeveloped or remote regions.

🧴 Dermatology

Using image classifier mobile apps or in a clinical environment, it is possible to evaluate:

•             Lesions on skin

•             Rashes

•             Possible melanomas

Advantage: Triage and Telemedicine The process of triage and remote dermatology let such as Telemedicine can take place.

Problems with Outsourcing the Classification of Medical Images

🔐 1. Information Security and Compliance

The exchange of patient data needs to be HIPAA, GDPR, and other health care-related regulations complied with.

Resolution: Only collaborate with companies, which adhere to the rules and ensure safe data delivery and the possibility of anonymizing medical image data.

🎯 2. Explainability and Transparency of models

The decisions made by AI in healthcare should be interpretable. Clinical decision-making does not agree with black-box algorithms.

Remedy: Select an AI provider which has explainable AI (XAI) capabilities i.e. heat maps, confidence scores or overlays to understand the segmentations.

🧪 3. Acceptance and Clinical Acceptance

Even the high-performance AI models should be checked on local datasets and in clinical practice to be accepted by the professionals.

Solution: prove the solution using the data of your institution and engage the clinicians in the validation and feedback.

🌐 4. Integrability with the World Today

The AI needs to be readily combined with EHRs, PACS systems or hospital processes.

Solution: Find work with vendors that support flexible APIs, HL7/DICOM compatibility and cloud/on-prem.

Best Healthcare-Providing Practices

Begin with a low-risk/high impact use case (e.g. triaging X-rays)

Take a pilot clinical to scale

•             Establish KPIs in terms of accuracy, turnaround time and false positive rates

•             Make a feedback system between the AI vendor and clinicians

•             Make sure that there are powerful data governance policies

But the Future: AI Clinical Assistant

Outsourcing in image classification is not intended to eliminate radiologists or pathologists, but to supplement them, to take some load off them, and to draw more attention to the cases with higher risk. With more specialized models becoming available, the AI becomes an effective clinical assistant to make quality care scalable to the population.

Indeed, the most innovative practices will be those who have outsourced their AI coupled with an in-house management, as the best of both worlds have been attained; that is, efficiency, affordability, and clinical accuracy.

In Conclusion

Image classification, object detection, object localization — all of these may be a tangled mess in your mind, and that's completely fine if you are new to these concepts. In reality, they are essential components of computer vision and image annotation, each with its own distinct nuances. Let's untangle the intricacies right away.We've already established that image classification refers to assigning a specific label to the entire image. On the other hand, object localization goes beyond classification and focuses on precisely identifying and localizing the main object or regions of interest in an image. By drawing bounding boxes around these objects, object localization provides detailed spatial information, allowing for more specific analysis.