CheXNet – a brief evaluation

CheXNet – a brief evaluation

Chest X-Ray deep dreamed - our AI & deep learning future
Chest Radiograph from ChestX-ray14 dataset processed with the deep dream algorithm trained on ImageNet

Andrew Ng released CheXNet yesterday on ArXiv (citation) and promoted it with a tweet which caused a bit of a stir on the internet and related radiology social media sites like Aunt Minnie.  Before Radiologists throw away their board certifications and look for jobs as Uber drivers, a few comments on what this does and does not do.

First off, from the Machine Learning perspective, methodologies check out.  It uses a 121 layer DenseNet, which is a powerful convolutional neural network.  While code has not yet been provided, the DenseNet seems similar to code repositories online where 121 layers are a pre-made format.  80/20 split for Training/Validation seems pretty reasonable (from my friend, Kirk Borne), Random initialization, minibatches of 16 w/oversampling positive classes, and a progressively decaying validation loss are utilized.  Class activation mappings are used to visualize areas in the image most indicative of the activated class (in this case, pneumonia).  This is an interesting technique that can be used to provide some human-interpretable insights into the potentially opaque DenseNet.

The last Fully Connected (FC) layer is replaced by a single output (only one class is being tested for – pneumonia) coupled to a sigmoid function (an activation function – see here) to give a probability between 0 and 1.

The test portion of the study was 420 Chest X-rays read by four radiologists, one of whom was a thoracic specialist.  They could choose between the 14 pathologies in the ChestX-ray14 dataset, read blind without any clinical data.

So, a ROC curve was created, showing three radiologists similar to each other, and one outlier.The radiologists lie slightly under the ROC curve of the CheXNet classifier.  But, a miss is as good as a mile, so the claims of at or above radiologist performance are accurate, because math.  Addendum – Even though this claim would likely not meet statistical significance.  Thanks, Luke.

So that’s the study.  Now, I will pick some bones with the study.

First, only including one thoracic radiologist is relevant, if you are going to make ground truth agreement of 3 out of four radiologists.  General radiologists will be less specific than specialist radiologists, and that is one of the reasons why we have moved to specialty-specific reads over the last 20 years.  If the three general rads disagreed with the thoracic rad, the thoracic rad’s ground truth would be discarded.  Think about this – you would take the word of the generalist over the specialist, despite greater training.  Google didn’t do this in their retinal machine learning paper.  Instead, Google used their three retinal specialists as ground truth and then looked at how the non-specialty opthalmologists were able to evaluate that data and what it meant to the training dataset.  (Thanks, Melody!)

Second, the Wang ChestXray14 dataset is a dataset that I believe was data-mined from NIH radiology reports.  This means that for the dataset, ground truth was whatever the radiologists said it was.  I’m not casting aspersions on the NIH radiologists, as I am sure they are pretty good.  I’m simply saying that the dataset’s ground truth is what it says it is, not necessarily what the patient’s clinical condition was.  As proof of that, here are a few cells from the findings field on this dataset.  I’m not sure if the multiple classes strengthens or weakens Mr. Ng’s argument.

Findings field from the ChestX-ray14 dataset (representative)

In any case, the NIH radiologists more than a few times perhaps couldn’t tell either, or identified one finding as the cause of the other (Infiltrate & Pneumonia mentioned side by side) and at the top you have an “atelectasis vs consolidation vs Pneumonia” (as radiologists we say these things).  Perhaps I am missing something here and the classifier is making a stronger decision between the pathologies.  But without the sigmoid activation percentages for each class in the 14, I can’t tell.  Andrew, if you read this, I have the utmost respect for you and your team, and I have learned from you.  But I would love to know your rebuttal, and I would urge you to publish those results.  Or perhaps someone should do it for reproducibility.

Addendum:   Other radiologists with machine learning chops whom I respect are also concerned about how the ground truth was decided in the ChestXRay14 dataset.  This needs to be further looked into.

Finally, I’m bringing up these points not to be a killjoy, but to be balanced.  I think it is important to see this and prevent an administrator from making a really boneheaded decision of firing their radiologists to put in a computer diagnostic system (not in the US, but elsewhere) and realizing it doesn’t work after spending a vast sum of money on it.  Startups competing in the field who do not have healthcare experience need to be aware of these pitfalls in their product.  I’m saying this because real people could be really hurt and impacted if we don’t manage this transition into AI well.

Thanks for reading, and feel free to comment here or on twitter or linkedin to me: @drsxr

Building a high-performance GPU computing workstation for deep learning – part I

This post is cross posted to .  For machine learning and AI issues, please visit the new site!

With Tensorflow released to the public, the NVidia Pascal Titan X GPU, along with (relatively) cheap storage and memory, the time was right to take the leap from CPU-based computing to GPU accelerated machine learning.

My venerable Xeon W3550 8GB T3500 running a 2GB Quadro 600 was outdated. Since a DGX-1 was out of the question ($129,000), I decided to follow other pioneers building their own deep learning workstations. I could have ended up with a multi-thousand dollar doorstop – fortunately, I did not.


  1. Reasonably fast CPU
  2. Current ‘Best’ NVidia GPU with large DDR5 memory
  3. Multi-GPU potential
  4. 32GB or more stable RAM
  5. SSD for OS
  6. Minimize internal bottlenecks
  7. Stable & Reliable – minimize hardware bugs
  8. Dual Boot Windows 10 Pro & Ubuntu 16.04LTS
  9. Can run: R, Rstudio, Pycharm, Python 3.5, Tensorflow


Total:                                                 $3725


Asus X99 E 10G WS Motherboard. Retail $699

A Motherboard sets the capabilities and configuration of your system. While newer Intel Skylake and Kaby Lake CPU architectures & chipsets beckon, reliability is important in a computationally intensive build, and their documented complex computation freeze bug makes me uneasy. Also, both architectures remain PCIe 3.0 at this time.

Therefore, I chose the ASUS X99 motherboard. The board implements 40 PCIe 3.0 lanes which will support three 16X PCIe 3.0 cards (i.e. GPU’s) and one 8x card. The PCIe 3.0-CPU lanes are the largest bottleneck in the system, so making these 16X helps the most.  It also has a 10G Ethernet jack somewhat future-proofing it as I anticipate using large datasets in the Terabyte size. It supports up to 128GB of DDR4. The previous versions of ASUS X99 WS have been well reviewed.


Intel Core i7 6850K Broadwell-E CPU Socket Retail $649

Socket LGA2011-v3 on the motherboard guides the CPU choice – the sweet spot in the Broadwell-E lineup is the overclockable 3.6Ghz 6850K with 6 cores and 15MB of L3 cache, permitting 40 PCIe lanes. $359 discounted is attractive compared to the 6900K, reviewed to offer minimal to no improvement at a $600 price premium. The 6950X is $1200 more for 4 extra cores, unnecessary for our purposes. Avoid the $650 6800K – pricier and slower with less (28) lanes. A stable overclock to 4.0Ghz is easily achievable on the 6850K.

NVidia GeForce 1080Ti 11GB – EVGA FTW3 edition Retail: $800

Last year, choosing a GPU was easy – the Titan X Pascal, a 12GB 3584 CUDA-core monster. However, by spring 2017 there were two choices: The Titan Xp, with slightly faster memory speed & internal bus, and 256 more CUDA cores; and the 1080Ti, the prosumer enthusiast version of the Titan X Pascal, with 3584 cores. The 1080Ti differs in its memory architecture – 11GB DDR5 and a slightly slower, slightly narrower bandwidth vs. the Xp.

The 1080Ti currently wins on price/performance. You can buy two 1080Ti’s for the price of one Titan Xp. Also, at time of purchase, Volta architecture was announced. As the PCIe bus is the bottleneck, and will remain so for a few years, batch size into DDR5 memory & CUDA cores will be where performance is gained. A 16GB DDR5 Volta processor would be a significant performance gain from a 12GB Pascal for deep learning. Conversely, 12GB Pascal to 11GB Pascal is a relative lesser performance hit. As I am later in the upgrade cycle, I’ll upgrade to the 16GB Volta and resell my 1080Ti in the future – I anticipate only taking a loss of $250 per 1080Ti on resell.

The FTW3 edition was chosen because it is a true 2-slot card (not 2.5) with better cooling than the Founder’s Edition 1080Ti. This will allow 3 to physically fit onto this motherboard.

64 GB DDR4-2666 DRAM – Corsair Vengeance low profile Retail : $600

DDR4 runs at 2133mhz unless overclocked. Attention must be paid to the size of the DRAM units to ensure they fit under the CPU cooler, which these do. From my research, DRAM speeds over 3000 lose stability. For Broadwell there’s not much evidence that speeds above 2666mhz improves performance. I chose 64GB because 1) I use R which is memory resident so the more GB the better and 2) There is a controversial rule of thumb that your RAM should equal 2x the size of your GPU memory to prevent bottlenecks. Implementing 3 1080Ti’s, 3x 11GB = 33 GB. Implementing 2 16GB Voltas would be 32GB.


Samsung 1TB 960 EVO M2 NVMe SSD Retail $500

The ASUS motherboard has a fast M2 interface, which, while using PCIe lanes, does not compete for slots or lanes. The 1TB size is sufficient for probably anything I will throw at it (all apps/programs, OS’s, and frequently used data and packages. Everything else can go on other storage. I was unnecessarily concerned about SSD heat throttling – on this motherboard, the slot’s location is in a good place which allows for great airflow over it. The speed in booting up Windows 10 or Ubuntu 16.04 LTS is noticeable.


EVGA Titanium 1200 power supply Retail $350

One of the more boring parts of the computer, but for a multi GPU build you need a strong 1200 or 1600W power supply. The high Titanium rating will both save on electricity and promote stability over long compute sessions.


Barracuda 8TB Hard Drive Retail $299

I like to control my data, so I’m still not wild about the cloud, although it is a necessity for very large data sets. So here is a large, cheap drive for on-site data storage. For an extra $260, I can Raid 1 the drive and sleep well at night.

Strike FUMA CPU Cooler. Retail $60

This was actually one of the hardest decisions in building the system – would the memory will fit under the fans? The answer is a firm yes. This dual fan tower cooler was well-rated, quiet, attractive, fit properly, half the price of other options, and my overclocked CPU runs extremely cool – 35C with full fan RPM’s, average operating temperature 42C and even under a high stress test, I have difficulty getting the temperature over 58C. Notably, the fans never even get to full speed on system control.


Corsair 750 D Airflow Edition Case. Retail $250

After hearing the horror stories of water leaks, I decided at this level of build not to go with water cooling. The 750D has plenty of space (enough for a server) for air circulation, and comes installed with 3 fans – two air intake on the front and one exhaust at upper rear. It is a really nice, sturdy, large case. My front panel was defective – the grating kept falling off – so Corsair shipped me a replacement quickly and without fuss.

Cougar Vortex 14” fans – Retail $20 ea.

Two extra cougar Vortex 14” fans were purchased, one as an intake fan at the bottom of the case, and one as a 2nd exhaust fan at the top of the case. These together create excellent airflow at noise levels I can barely hear. Two fans on the CPU Heat Sink plus Three Fans on the GPU plus five fans on the case plus one in the power supply = 11 fans total! More airflow at lower RPM = silence.


Windows 10 Pro USB edition Retail $199

This is a dual boot system so, there you go.

Specific limitations with this system are as follows. While it will accept four GPU’s physically, the slots are limited to 16X/16X/16X/8X with the M2 drive installed which may affect performance on the 4th GPU (& therefore deep learning model training and performance). Additionally, the CPU upgrade path is limited – without going to a Xeon, the only reasonable upgrade from the 6850K’s 14,378 passmark is the 6950X, with a passmark of 20,021. In the future if more than 128GB DDR4 is required, that will be a problem with this build.

Finally, inherent bandwidth limitations exist in the PCIe 3.0 protocol and aren’t easily circumvented. PCIe 3.0 throughput is 8GB/s. Compare this to NVidia’s proprietary NVlink that allows throughput of 20-25GB/s (Pascal vs. Volta). Note that current NVlink speeds will not be surpassed until PCIe5.0 is implemented at 32GB/s in 2019. NVidia’s CUDA doesn’t implement SLI, either, so at present that is not a solution. PCIe 4.0 has just been released with only IBM adopting, doubling transfer vs. 3.0, and 5.0 has been proposed, doubling yet again. However, these faster protocols may be difficult and/or expensive to implement. A 4 slot PCIe 5.0 bus will probably not be seen until into the 2020’s. This means that for now, dedicated NVlink 2.0 systems will outperform similar PCIe systems.

With that said, this system approaches a best possible build considering price and reliability, and should be able to give a few years of good service, especially if the GPU’s are upgraded periodically. Precursor systems based upon the Z97 chipset are still viable for deep learning, albeit with slower speeds, and have been matched to older NVidia 8GB 1070 GPU’s which are again half the price of the 1080Ti.

In part II, I will describe how I set up the system configuration for dual boot and configured deep learning with Ubuntu 16.04LTS. Surprisingly, this was far more difficult than the actual build itself, for multiple reasons I will explain & detail with the solutions.  And yes, it booted up.  On the first try.

Machine Intelligence in Medical Imaging Conference – Report

blueI heard about the Society of Imaging Informatics in Medicine’s (SIIM) Scientific Conference on Machine Intelligence in Medical Imaging (C-MIMI) on Twitter.  Priced attractively, easy to get to, I’m interested in Machine Learning and it was the first radiology conference I’ve seen on this subject, so I went.  Organized on short notice so I was expecting a smaller conference.


I almost didn’t get a seat.  It was packed.

The conference had real nuts and bolts presentations & discussions on healthcare imaging machine learning (ML).  Typically, these were Convolutional Neural Networks (CNN‘s/Convnets) but a few Random Forests (RF) and Support Vector Machines (SVM) sneaked in, particularly in hybrid models along with a CNN (c.f.  Microsoft).  Following comments assume some facility in understanding/working with Convnets.

Some consistent threads throughout the conference:

  • Most CNN’s were trained on Imagenet with the final fully connected (FC) layer removed; then re-trained on radiology data with a new classifer FC layer placed at the end.
  • Most CNN’s were using Imagenet standard three layer RGB input despite being greyscale.  This is of uncertain significance and importance.
  • The limiting of input matrices to grids less than image size is inherited from the Imagenet competitions (and legacy computational power).  Decreased resolution is a limiting factor in medical imaging applications, potentially worked-around by multi-scale CNN’s.
  • There is no central data repository for a good “Ground Truth” to develop improved machine imaging models.
  • Data augmentation methods are commonly used due to lower numbers of obtained cases.

Keith Dryer DO PhD gave an excellent lecture about the trajectory of machine imaging and how it will be an incremental process with AI growth more narrow in scope than projected, chiefly limited by applications.  At this time, CNN creation and investigation is principally an artisanal product with limited scalability.  There was a theme – “What is ground truth?” which in different instances is different things (path proven, followed through time, pathognomonic imaging appearance).

There was an excellent educational session from the FDA’s Berkman Sahiner.  The difference between certifying a type II or type III device may keep radiologists working longer than expected!  A type II device, like CAD, identifies a potential abnormality but does not make a treatment recommendation and therefore only requires a 510(k) application.  A type III device, as in an automated interpretation program creating diagnosis and treatment recommendations will require a more extensive application including clinical trials, and a new validation for any material changes.  One important insight (there were many) was that the FDA requires training and test data to be kept separate.   I believe this means that simple cross-validation is not acceptable nor sufficient for FDA approval or certification.  Adaptive systems may be a particularly challenging area for regulation, as similar to the ONC, significant changes to the software of the algorithm will require a new certification/approval process.

Industry papers were presented from HK Lau of Arterys, Xiang Zhou of Siemens, Xia Li of GE, and Eldad Elnekave of Zebra medical.  The Zebra medical presentation was impressive, citing their use of the Google Inception V3 model and a false-color contrast limited adaptive histogram equalization algorithm, which not only provides high image contrast with low noise, but also gets around the 3-channel RGB issue.  Given statistics for their CAD program were impressive at 94% accuracy compared to a radiologist at 89% accuracy.

Scientific Papers were presented by Matthew Chen, Stanford; Synho Do, Harvard; Curtis Langlotz, Stanford; David Golan, Stanford; Paras Lakhani, Thomas Jefferson; Panagiotis Korfiatis, Mayo Clinic; Zeynettin Akkus, Mayo Clinic; Etka Bullar, U Saskatchewan; Mahmudur Rahman, Morgan State U; Kent Ogden SUNY upstate.

Ronald Summers, MD PhD from the NIH gave a presentation on the work from his lab in conjunction with Holger Roth, detailing the specific CNN approaches to Lymph Node detection, Anatomic level detection, Vertebral body segmentation, Pancreas Segmentation, and colon polyp screening with CT-colonography, which had high False Positives.  In his experience, deeper models performed better.  His lab also changes unstructured radiology reporting into structured reporting through ML techniques.

Abdul Halabi of NVIDIA gave an impressive presentation on the supercomputer-like DGX-1 GPU cluster (5 deliveries to date, the fifth of which was to Mass. General, a steal at over $100K), and the new Pascal architecture in the P4 & P40 GPU’s.  60X performance on AlexNet vs the original version/GPU configuration in 2012.  Very impressive.

Sayan Pathak of Microsoft Research and the Inner Eye team gave a good presentation where he demonstrated that a RF was really just a 2 layer DNN, i.e. a sparse 2 layer perceptron.   Combining this with a CNN (dNDE.NET), it beat googLENet’s latest version in the Imagenet arms race.  However, as one needs to solve for both structures simultaneously, it is an expensive (long, intense) computation.

Closing points were the following:

  • Most devs currently using Python – Tensorflow +/- Keras with fewer using CAFFE off of  Modelzoo
  • De-identification of data is a problem, even moreso when considering longitudinal followup.
  • Matching accuracy to the radiologist’s report may not be as important as actual outcomes report.
  • There was a lot of interest in organizing a competition to advance medical imaging, c.f. Kaggle.
  • Radiologists aren’t obsolete just yet.

It was a great conference.  An unexpected delight.  Food for your head!




Memory requirements for Convolutional Neural Network analysis of brain MRI.

Believed to be in the publc domainI’m auditing the wonderful Stanford CS 231n class on Convolutional Neural Networks in Computer Vision.

A discussion the other day was on the amount of memory required to analyze one image as it goes through the Convolutional Neural Network (CNN). This was interesting – how practical is it for application to radiology imaging?  (To review some related concepts see my earlier post : What Big Data  Visualization Analytics can learn from Radiology)

Take your standard non-contrast MRI of the brain. There are 5 sequences (T1, T2, FLAIR, DWI, ADC). For the purposes of this analysis, all axial. Assume a 320×320 viewing matrix for each slice. Therefore, one image will be a 320x320x5 matrix suitable for processing into a 512,000 byte vector. Applying this to the VGGNet Protocol D (1) yields the following:


In each image, there are 320 x and y pixels and each pixel holding a greyscale value. There are 5 different sequences. Each axial slice takes up 512KB, the first convolutional layers hold most of the memory at 6.4MB each, and summing all layers uses 30.5MB. Remember that you have to double the memory for the forward/backward pass through the network, giving you 61MB per image. Finally, the images do not exist in a void, but are part of about 15 axial slices of the head, giving you a memory requirement of 916.5MB, or about a gigabyte.

Of course, that’s just for feeding an image through the algorithm.

This is simplistic because:

  1. VGG is not going to get you to nearly enough accuracy for diagnosis! (50% accurate, I’m guessing)
  2. The MRI data is only put into slices for people to interpret – the data itself exists in K-space. What that would do to machine learning interpretation is another discussion.
  3. We haven’t even discussed speed of training the network.
  4. This is for older MRI protocols.  Newer MRI’s have larger matrices (512×512) and thinner slices (3mm) available, which will increase the necessary memory to approximately 4GB.

Nevertheless, it is interesting to note that the amount of memory required to train a neural network of brain MRI’s is in reach of a home network enthusiast.

(1). Karen Simonyan & Andrew Zisserman, Very Deep Convolutional Networks for Large-Scale Visual Recognition, ICLR 2015

The coming computer vision revolution

3 layer (7,5,3 hidden layers) neural network created in R using the neuralnet package.
3 layer (7,5,3 hidden layers) neural network created in R using the neuralnet package.


Nothing of him that doth fade
But doth suffer a sea-change
Into something rich and strange.

– Shakespeare, The Tempest 1.2.396-401

I’m halfway through auditing Stanford’s CS231n course – Convolutional Neural Networks for Visual Recognition.

Wow. Just Wow. There is a sea-changing paradigm shift that is happening NOW –  we probably have not fully realized it yet.

We are all tangentially aware of CV applications in our daily lives – Facebook’s ability to find us in photos, optical character recognition (OCR) of our address on postal mail, that sort of thing. But these algorithms were rule-based expert systems grounded in supervised learning methods. Applications were largely one-off for a specific, single task. They were expensive, complicated, and somewhat error prone.

So what changed?   First, a little history. In the early 1980’s I had a good friend obtaining a MS in comp sci all atwitter about “Neural Networks”. Back then they went nowhere. Too much processing/memory/storage required, too difficult to tune, computationally slow. Fail.


1999 –  Models beginning with SIFT & ending with SVM (support vector machine) deformable parts. Best model only 74% accurate.

2006 – Restricted Boltzmann Machines apply backpropogation to allow deep neural networks.

2012 – AlexNet Deep learning applied to Imagenet classification database competition achieves a nearly 2X increase in accuracy to earlier SVM methods.

2015-   ResNet Deep learning system achieves a 4.5X increase in accuracy compared to Alexnet and 8X increase in accuracy to old SVM models.

In practical aspects, what does this mean? On a data set with 1000 different items  (ImageNet), ResNet is getting the item 100% correct (compared to a human) about 80% of the time, and correctly classifies the image as one of a list of 5 most probable items 96.4% of the time. People are typically believed to have 95% accuracy identifying the correct image. It’s clear to see that the computer is not far off.

2012 was the watershed year with the first application and win of the CNN to the dataset, and the improvement was significant enough it sparked additional refinements and development. That is still going on – the ResNet example was just released in December 2015! It’s clear that this is an area of active research and further improvements expected.

The convolutional neural network is a game-changer and will likely approach and perhaps exceed human accuracy in computer vision and classification in the near future. That’s a big deal.  As this is a medical blog, the applications to healthcare are obvious – radiology, pathology, dermatology, ophthalmology for starters.  But the CNN may also be useful for the complicated process problems I’ve developed here on the blog – the flows themselves resemble networks naturally.  So why not model them as such?  Why is it a game changer?  Because the model is probably universally adaptable to visual classification problems and once trained, potentially cheap.


I’ll  write more on this in the coming weeks – I’ve been inching towards deep learning models (but lagging blogging about them) but there is no reason to wait any more. The era of the deep learning neural network is here.